{"id":19155,"date":"2026-09-19T16:19:08","date_gmt":"2026-09-19T14:19:08","guid":{"rendered":"https:\/\/www.beseit.net\/?p=19155"},"modified":"2026-09-19T16:20:24","modified_gmt":"2026-09-19T14:20:24","slug":"servidor-local-tts-server-i-piper-ona-a-deepin-corregit","status":"publish","type":"post","link":"https:\/\/www.beseit.net\/?p=19155","title":{"rendered":"servidor local tts-server i Piper (ona) a Deepin corregit"},"content":{"rendered":"\n

UOS AI<\/p>\n\n\n\n

iquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho \\\u00bbHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | aplay -f S16_LE -r 22050 -c 1 Playing WAVE ‘stdin’ : Signed 16 bit Little Endian, Rate 22050 Hz, Mono Miquel@Miquel-PC:~$ aquest issatge es reprodueix sense problemes amb el servidor local tts-server i Piper (ona) a Deepin. podries mostra el flux del texte fins arribar al altaveus?<\/p>\n\n\n\n

Deep think completado (tard\u00f3 3s)<\/p>\n\n\n\n

Aqu\u00ed tens el flux complet del text fins als altaveus, pas a pas:<\/p>\n\n\n\n

Flux del text \u2192 so<\/h2>\n\n\n\n

plaintext<\/p>\n\n\n\n

1TEXT (stdin)<\/p>\n\n\n\n

2 \u2502<\/p>\n\n\n\n

3 \u25bc<\/p>\n\n\n\n

4\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<\/p>\n\n\n\n

5\u2502 docker exec piper bash -c \u00ab…\u00bb \u2502<\/p>\n\n\n\n

6\u2502 \u2502 \u2502<\/p>\n\n\n\n

7\u2502 \u25bc \u2502<\/p>\n\n\n\n

8\u2502 echo \u00abHola, s\u00f3c l’Ona…\u00bb \u2500\u2500\u25ba stdin \u2502<\/p>\n\n\n\n

9\u2502 \u2502 \u2502<\/p>\n\n\n\n

10\u2502 \u25bc \u2502<\/p>\n\n\n\n

11\u2502 piper –model ca_ES-upc_ona-medium.onnx \u2502<\/p>\n\n\n\n

12\u2502 1. Normalitza el text (xifres, abreviatures…) \u2502<\/p>\n\n\n\n

13\u2502 2. Fonetitza (text \u2192 fonemes) \u2502<\/p>\n\n\n\n

14\u2502 3. S\u00edntesi neuronal (fonemes \u2192 mel-espectrograma) \u2502<\/p>\n\n\n\n

15\u2502 4. Vocoder (mel-espectrograma \u2192 mostres PCM) \u2502<\/p>\n\n\n\n

16\u2502 \u2502 \u2502<\/p>\n\n\n\n

17\u2502 \u25bc \u2502<\/p>\n\n\n\n

18\u2502 Sortida: PCM cru (S16_LE, 22050 Hz, mono) \u2500\u2500\u25ba stdout \u2502<\/p>\n\n\n\n

19\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n\n\n\n

20 \u2502<\/p>\n\n\n\n

21 \u2502 pipe (|) \u2014 dins de l’host, fora del container<\/p>\n\n\n\n

22 \u25bc<\/p>\n\n\n\n

23\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<\/p>\n\n\n\n

24\u2502 aplay -f S16_LE -r 22050 -c 1 \u2502<\/p>\n\n\n\n

25\u2502 \u2502 \u2502<\/p>\n\n\n\n

26\u2502 \u25bc \u2502<\/p>\n\n\n\n

27\u2502 ALSA (Advanced Linux Sound Architecture) \u2502<\/p>\n\n\n\n

28\u2502 \u2502 \u2502<\/p>\n\n\n\n

29\u2502 \u25bc \u2502<\/p>\n\n\n\n

30\u2502 Controlador del so (driver) \u2502<\/p>\n\n\n\n

31\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n\n\n\n

32 \u2502<\/p>\n\n\n\n

33 \u25bc<\/p>\n\n\n\n

34DAC (convertidor digital\u2192anal\u00f2gic) de la targeta de so<\/p>\n\n\n\n

35 \u2502<\/p>\n\n\n\n

36 \u25bc<\/p>\n\n\n\n

37Amplificador \u2192 Altaveus \ud83d\udd0a<\/p>\n\n\n\n

38<\/p>\n\n\n\n

38 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

Detalls clau de cada etapa<\/h2>\n\n\n\n
    \n
  1. docker exec<\/code><\/strong> \u2014 Executa el proc\u00e9s dins del contenidor piper<\/code>, que cont\u00e9 el model i l’eina.<\/li>\n\n\n\n
  2. echo ... | piper<\/code><\/strong> \u2014 El text entra per stdin<\/em>. Piper fa la s\u00edntesi amb el model neuronal ONNX ca_ES-upc_ona-medium<\/code>:\n
      \n
    • El model converteix el text en una seq\u00fc\u00e8ncia de fonemes (representaci\u00f3 interm\u00e8dia).<\/li>\n\n\n\n
    • La xarxa neuronal genera l’espectrograma mel (representaci\u00f3 freq\u00fcencial de la veu).<\/li>\n\n\n\n
    • El vocoder<\/strong> converteix aquest espectrograma en mostres d’\u00e0udio PCM.<\/li>\n<\/ul>\n<\/li>\n\n\n\n
    • --output_file -<\/code><\/strong> \u2014 L’\u00e0udio es treu per stdout<\/em> en format PCM cru<\/strong> (sense cap\u00e7alera WAV), a 22050 Hz, mono, 16 bits little-endian.<\/li>\n\n\n\n
    • | aplay<\/code><\/strong> \u2014 La canonada passa aquest flux de dades a aplay<\/code> de l’host:\n
        \n
      • -f S16_LE<\/code> \u2192 interpreta les mostres com a enter amb signe de 16 bits little-endian.<\/li>\n\n\n\n
      • -r 22050<\/code> \u2192 freq\u00fc\u00e8ncia de mostreig.<\/li>\n\n\n\n
      • -c 1<\/code> \u2192 un canal (mono).<\/li>\n<\/ul>\n<\/li>\n\n\n\n
      • ALSA \u2192 DAC \u2192 Altaveus<\/strong> \u2014 aplay<\/code> entrega les mostres al subsistema de so del kernel; la targeta de so converteix les mostres digitals en senyal anal\u00f2gic, que l’amplificador fa arribar als altaveus.<\/li>\n<\/ol>\n\n\n\n

        Observaci\u00f3<\/h2>\n\n\n\n

        El fet que funcioni aix\u00ed demostra que el contenidor genera \u00e0udio PCM cru a stdout<\/strong> i que el pipe el porta directament a l’ALSA de Deepin sense cap intermediari (sense PulseAudio\/PipeWire pel mig), cosa que minimitza la lat\u00e8ncia. Si volguessis enregistrar-ho en un fitxer WAV, n’hi hauria prou amb afegir-hi una cap\u00e7alera:<\/p>\n\n\n\n

        bash<\/p>\n\n\n\n

        1docker exec piper bash -c \u00abecho \\\u00bb…\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | sox -t raw -r 22050 -e signed -b 16 -c 1 – output.wav<\/p>\n\n\n\n

        2<\/p>\n\n\n\n

        2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

        Vols que t’expliqui cap etapa amb m\u00e9s profunditat?<\/p>\n\n\n\n

        Enrutamiento Inteligente<\/p>\n\n\n\n

        TEXT (stdin) qu\u00e8 vol dir?<\/p>\n\n\n\n

        Deep think completado (tard\u00f3 2s)<\/p>\n\n\n\n

        TEXT (stdin)<\/code><\/strong> vol dir que el text arriba al programa piper<\/strong> per l’entrada est\u00e0ndard<\/strong> (standard input<\/em>), no pas com a argument ni com a fitxer.<\/p>\n\n\n\n

        Qu\u00e8 \u00e9s stdin?<\/h2>\n\n\n\n

        Cada proc\u00e9s a Linux t\u00e9 tres canals de comunicaci\u00f3 per defecte:<\/p>\n\n\n\n

        Canal<\/th>Nom<\/th>Funci\u00f3<\/th><\/tr><\/thead>
        stdin<\/strong> (0)<\/td>Entrada est\u00e0ndard<\/td>D’on el programa llegeix<\/strong> dades<\/td><\/tr>
        stdout (1)<\/td>Sortida est\u00e0ndard<\/td>On el programa escriu<\/strong> resultats<\/td><\/tr>
        stderr (2)<\/td>Error est\u00e0ndard<\/td>On escriu missatges d’error<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

        Quan executes un programa des del terminal i no li dius d’on treure les dades, llegeix de stdin<\/strong> (que normalment \u00e9s el teclat).<\/p>\n\n\n\n

        Al teu cas concret<\/h2>\n\n\n\n

        bash<\/p>\n\n\n\n

        1echo \u00abHola, s\u00f3c l’Ona…\u00bb | piper –model … –output_file –<\/p>\n\n\n\n

        2<\/p>\n\n\n\n

        2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

        Aqu\u00ed passa aix\u00f2:<\/p>\n\n\n\n

        plaintext<\/p>\n\n\n\n

        1echo \u00abHola…\u00bb pipe (|) piper<\/p>\n\n\n\n

        2 \u2502 \u2502<\/p>\n\n\n\n

        3 escriu el text \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u25ba llegeix de stdin<\/p>\n\n\n\n

        4<\/p>\n\n\n\n

        4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

          \n
        1. echo<\/code> escriu<\/strong> el text al seu stdout.<\/li>\n\n\n\n
        2. La canonada<\/strong> (|<\/code>) connecta el stdout de echo<\/code> amb el stdin<\/strong> de piper<\/code>.<\/li>\n\n\n\n
        3. piper<\/code> llegeix<\/strong> el text del seu stdin (com si vingu\u00e9s del teclat, per\u00f2 ve del pipe) i el processa.<\/li>\n<\/ol>\n\n\n\n

          \u00c9s per aix\u00f2 que el diagrama deia TEXT (stdin)<\/strong>: el text entra a piper per aquest canal est\u00e0ndard. Com que no especifiques cap fitxer d’entrada (--input_file<\/code>), piper agafa les dades de stdin.<\/p>\n\n\n\n

          Per qu\u00e8 s’escriu aix\u00ed?<\/h2>\n\n\n\n
            \n
          • |<\/code> (pipe)<\/strong> \u2192 encadena processos: la sortida d’un \u00e9s l’entrada del seg\u00fcent.<\/li>\n\n\n\n
          • --output_file -<\/code><\/strong> \u2192 el guionet -<\/code> \u00e9s una convenci\u00f3 que significa \u00abel mateix stdin\/stdout en lloc d’un fitxer\u00bb, \u00e9s a dir, escriu l’\u00e0udio al stdout<\/em> perqu\u00e8 despr\u00e9s aplay<\/code> el pugui rebre.<\/li>\n<\/ul>\n\n\n\n

            Resumint: stdin<\/strong> \u00e9s simplement \u00abel lloc per on piper rep el text d’entrada\u00bb \u2014 en aquest cas, a trav\u00e9s del pipe des de echo<\/code>.<\/p>\n\n\n\n

            Enrutamiento Inteligente<\/p>\n\n\n\n

            tic el problema que l’entrada nom\u00e8s es llegeix si el missat t\u00e9 un minin de caracters, per example ‘teulada’ es perdria. qui pot ser el responsable?<\/p>\n\n\n\n

            Deep think completado (tard\u00f3 30s)<\/p>\n\n\n\n

            Molt bona pregunta. La resposta curta: el sospit\u00f3s principal \u00e9s aplay<\/code><\/strong> (o m\u00e9s exactament, com es gestiona el final del flux quan es llegeix d’un pipe<\/em>), no pas Piper<\/strong>.<\/p>\n\n\n\n

            Per qu\u00e8?<\/h2>\n\n\n\n

            Amb un text curt com teulada<\/code>, Piper genera l’\u00e0udio correctament, per\u00f2 l’\u00e0udio \u00e9s tan breu que tot el contingut pot quedar dins el buffer de la targeta de so<\/strong> sense arribar a sonar. La seq\u00fc\u00e8ncia \u00e9s:<\/p>\n\n\n\n

            plaintext<\/p>\n\n\n\n

            1piper genera poques dades \u2500\u2500\u25ba pipe \u2500\u2500\u25ba aplay escriu al buffer de ALSA<\/p>\n\n\n\n

            2 \u2502<\/p>\n\n\n\n

            3 EOF (el contenidor acaba) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524<\/p>\n\n\n\n

            4 \u25bc<\/p>\n\n\n\n

            5 aplay tanca el dispositiu de so<\/p>\n\n\n\n

            6 (snd_pcm_close) i descarta el que queda<\/p>\n\n\n\n

            7<\/p>\n\n\n\n

            7 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

            Quan aplay<\/code> rep EOF<\/strong> pel stdin (perqu\u00e8 el contenidor ha acabat i el pipe es tanca), tanca el dispositiu de so immediatament<\/strong>, sovint sense esperar que el buffer de la targeta es buidi del tot. Amb missatges llargs, nom\u00e9s es talla un tros final petit i no ho notes; amb teulada<\/code>, l’\u00e0udio sencer pot estar al buffer i es perd.<\/p>\n\n\n\n

            Com confirmar-ho (diagn\u00f2stic r\u00e0pid)<\/h2>\n\n\n\n

            Prova de guardar la sortida en un fitxer i reproduir-lo despr\u00e9s:<\/p>\n\n\n\n

            bash<\/p>\n\n\n\n

            1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb > \/tmp\/teulada.wav<\/p>\n\n\n\n

            2aplay \/tmp\/teulada.wav<\/p>\n\n\n\n

            3<\/p>\n\n\n\n

            3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              \n
            • Si sona b\u00e9<\/strong> des del fitxer \u2192 Piper funciona; el problema \u00e9s el pipe + aplay<\/strong>.<\/li>\n\n\n\n
            • Si el fitxer \u00e9s buit o gaireb\u00e9 buit<\/strong> \u2192 llavors seria Piper (menys probable, per\u00f2 ho descartem).<\/li>\n<\/ul>\n\n\n\n

              Solucions pr\u00e0ctiques<\/h2>\n\n\n\n

              Opci\u00f3 1 \u2014 Afegir un coix\u00ed de silenci<\/strong> al final, perqu\u00e8 aplay tingui m\u00e9s dades i nom\u00e9s es talli el silenci:<\/p>\n\n\n\n

              bash<\/p>\n\n\n\n

              1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb \\<\/p>\n\n\n\n

              2 | sox -t wav – -t wav – pad 0 0.5 | aplay<\/p>\n\n\n\n

              3<\/p>\n\n\n\n

              3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              Opci\u00f3 2 \u2014 Reproduir amb el servidor de so<\/strong> (PipeWire\/PulseAudio), que s\u00ed que drena el buffer correctament abans de tancar:<\/p>\n\n\n\n

              bash<\/p>\n\n\n\n

              1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model … –output_file -\u00bb | pw-play –<\/p>\n\n\n\n

              2# o b\u00e9:<\/p>\n\n\n\n

              3docker exec piper bash -c \u00abecho ‘teulada’ | piper –model … –output_file -\u00bb | paplay –<\/p>\n\n\n\n

              4<\/p>\n\n\n\n

              4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              Opci\u00f3 3 \u2014 Fitxer temporal<\/strong> (la m\u00e9s robusta):<\/p>\n\n\n\n

              bash<\/p>\n\n\n\n

              1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model … –output_file -\u00bb > \/tmp\/veu.wav<\/p>\n\n\n\n

              2aplay \/tmp\/veu.wav<\/p>\n\n\n\n

              3<\/p>\n\n\n\n

              3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              Opci\u00f3 4 \u2014 Si vols mantenir el pipe amb aplay<\/strong>, for\u00e7a un drenatge expl\u00edcit amb un petit retard final:<\/p>\n\n\n\n

              bash<\/p>\n\n\n\n

              1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model … –output_file -; sleep 1\u00bb | aplay …<\/p>\n\n\n\n

              2<\/p>\n\n\n\n

              2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              (el sleep<\/code> dins el contenidor fa que el pipe no es tanqui tan de pressa).<\/p>\n\n\n\n

              \n\n\n\n

              Prova primer el diagn\u00f2stic amb fitxer<\/strong> per confirmar-ho. Si vols, puc ajudar-te a muntar un script que gestioni autom\u00e0ticament el cas dels textos curts. Qu\u00e8 prefereixes?<\/p>\n\n\n\n

              Enrutamiento Inteligente<\/p>\n\n\n\n

              Miquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho ‘teulada’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb > \/tmp\/teulada.wav aplay \/tmp\/teulada.wav Playing raw data ‘\/tmp\/teulada.wav’ : Unsigned 8 bit, Rate 8000 Hz, Mono Miquel@Miquel-PC:~$<\/p>\n\n\n\n

              Deep think completado (tard\u00f3 27s)<\/p>\n\n\n\n

              Molt revelador! Aquest missatge d’aplay \u00e9s la clau del misteri:<\/p>\n\n\n\n

              plaintext<\/p>\n\n\n\n

              1Playing raw data ‘\/tmp\/teulada.wav’ : Unsigned 8 bit, Rate 8000 Hz, Mono<\/p>\n\n\n\n

              2<\/p>\n\n\n\n

              2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              Qu\u00e8 vol dir aquest missatge?<\/h2>\n\n\n\n

              aplay diu \u00abraw data\u00bb<\/strong> amb format per defecte (U8, 8000 Hz, mono) \u2014 aix\u00f2 \u00e9s el que fa quan no troba cap cap\u00e7alera WAV<\/strong> al fitxer. \u00c9s a dir, el fitxer \/tmp\/teulada.wav<\/code> \u00e9s buit o gaireb\u00e9 buit<\/strong> (menys de 44 bytes).<\/p>\n\n\n\n

              Comparem amb el primer comandament, on deia:<\/p>\n\n\n\n

              plaintext<\/p>\n\n\n\n

              1Playing WAVE ‘stdin’ : Signed 16 bit Little Endian, Rate 22050 Hz, Mono<\/p>\n\n\n\n

              2<\/p>\n\n\n\n

              2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

              All\u00e0 aplay s\u00ed<\/strong> va detectar la cap\u00e7alera WAV. La difer\u00e8ncia no \u00e9s aplay: \u00e9s que Piper no ha generat l’\u00e0udio per a \u00abteulada\u00bb<\/strong>.<\/p>\n\n\n\n

              El responsable: la segmentaci\u00f3 de frases de Piper<\/h2>\n\n\n\n

              Piper nom\u00e9s sintetitza frases completes<\/strong>. Divideix el text en frases buscant puntuaci\u00f3 de final de frase (.<\/code>, !<\/code>, ?<\/code>):<\/p>\n\n\n\n

                \n
              • \"Hola, s\u00f3c l'Ona... Barcelona.\"<\/code> \u2192 acaba amb punt \u2192 frase completa \u2192 sintetitza<\/strong> \u2705<\/li>\n\n\n\n
              • \"teulada\"<\/code> \u2192 no acaba amb cap puntuaci\u00f3<\/strong> \u2192 l’\u00faltim fragment es considera incomplet i es descarta<\/strong> a l’EOF \u274c<\/li>\n<\/ul>\n\n\n\n

                \u00c9s un comportament conegut de Piper: si el text no acaba amb .<\/code>, !<\/code> o ?<\/code>, no produeix cap sortida<\/strong>.<\/p>\n\n\n\n

                Comprovaci\u00f3 r\u00e0pida<\/h2>\n\n\n\n

                bash<\/p>\n\n\n\n

                1ls -l \/tmp\/teulada.wav # espera 0 bytes o < 44<\/p>\n\n\n\n

                2<\/p>\n\n\n\n

                2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                bash<\/p>\n\n\n\n

                1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb > \/tmp\/teulada2.wav<\/p>\n\n\n\n

                2aplay \/tmp\/teulada2.wav # ara hauria de dir \u00abPlaying WAVE…\u00bb i sonar<\/p>\n\n\n\n

                3<\/p>\n\n\n\n

                3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                Soluci\u00f3<\/h2>\n\n\n\n

                Assegura que el text sempre acabi amb puntuaci\u00f3<\/strong> abans de passar-lo a Piper. Per exemple:<\/p>\n\n\n\n

                bash<\/p>\n\n\n\n

                1docker exec piper bash -c \u00abecho \\\u00bbteulada.\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | aplay -f S16_LE -r 22050 -c 1<\/p>\n\n\n\n

                2<\/p>\n\n\n\n

                2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                Si ho fas des d’un script amb text variable, pots normalitzar-ho aix\u00ed:<\/p>\n\n\n\n

                bash<\/p>\n\n\n\n

                1TEXT=\u00bbteulada\u00bb<\/p>\n\n\n\n

                2# Afegeix punt si el text no acaba en puntuaci\u00f3<\/p>\n\n\n\n

                3[[ \u00ab$TEXT\u00bb =~ [.!?]$ ]] || TEXT=\u00bb$TEXT.\u00bb<\/p>\n\n\n\n

                4<\/p>\n\n\n\n

                5docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | aplay -f S16_LE -r 22050 -c 1<\/p>\n\n\n\n

                6<\/p>\n\n\n\n

                6 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                Prova el test de teulada.<\/code> amb punt i digue’m si ja sona \u2014 si \u00e9s aix\u00ed, el problema era exactament aquest.<\/p>\n\n\n\n

                Enrutamiento Inteligente<\/p>\n\n\n\n

                Miquel@Miquel-PC:~$ ls -l \/tmp\/teulada.wav # espera 0 bytes o < 44 -rw-rw-r– 1 Miquel Miquel 0 de set. 19 11:14 \/tmp\/teulada.wav Miquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb > \/tmp\/teulada2.wav aplay \/tmp\/teulada2.wav # ara hauria de dir \u00abPlaying WAVE…\u00bb i sonar Playing raw data ‘\/tmp\/teulada2.wav’ : Unsigned 8 bit, Rate 8000 Hz, Mono Miquel@Miquel-PC:~$ no ha sonat encara que els altaveus han fet un sorollet?<\/p>\n\n\n\n

                Deep think completado (tard\u00f3 41s)<\/p>\n\n\n\n

                Aquest resultat descarta la hip\u00f2tesi de la puntuaci\u00f3: fins i tot amb teulada.<\/code> no hi ha sortida. El \u00absorollet\u00bb<\/strong> que has sentit \u00e9s simplement el clic de la targeta de so en obrir-se i tancar-se<\/strong> sense rebre dades \u2014 aplay no va reproduir res perqu\u00e8 el fitxer era buit (o quasi).<\/p>\n\n\n\n

                On som ara<\/h2>\n\n\n\n
                Prova<\/th>Resultat<\/th>Conclusi\u00f3<\/th><\/tr><\/thead>
                teulada<\/code> (sense punt)<\/td>0 bytes<\/td>Piper no genera res<\/td><\/tr>
                teulada.<\/code> (amb punt)<\/td>\u00abraw data\u00bb \u2192 buit<\/td>La puntuaci\u00f3 no \u00e9s la causa<\/strong><\/td><\/tr>
                Missatge llarg original<\/td>\u00abPlaying WAVE\u00bb \u2705<\/td>Piper funciona amb text llarg<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                El sospit\u00f3s ha canviat: Piper no genera cap sortida per a textos curts<\/strong> (o b\u00e9 docker exec perd la sortida petita). Cal separar aquestes dues possibilitats.<\/p>\n\n\n\n

                Diagn\u00f2stic 1: Quant surt exactament per stdout?<\/h2>\n\n\n\n

                bash<\/p>\n\n\n\n

                1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | wc -c<\/p>\n\n\n\n

                2<\/p>\n\n\n\n

                2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                  \n
                • Si d\u00f3na 0<\/strong> \u2192 Piper no sintetitza res (el problema \u00e9s dins el contenidor).<\/li>\n\n\n\n
                • Si d\u00f3na > 0<\/strong> \u2192 Piper s\u00ed que genera \u00e0udio, per\u00f2 el redireccionament\/docker exec el perd (problema de flux).<\/li>\n<\/ul>\n\n\n\n

                  Diagn\u00f2stic 2: Prova dins del contenidor (evita el stdout de docker)<\/h2>\n\n\n\n

                  bash<\/p>\n\n\n\n

                  1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/test.wav && wc -c \/tmp\/test.wav\u00bb<\/p>\n\n\n\n

                  2<\/p>\n\n\n\n

                  2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                    \n
                  • Si el fitxer dins el contenidor t\u00e9 > 0 bytes<\/strong> \u2192 Piper funciona; el problema \u00e9s el stdout de docker exec<\/strong> (buffering\/truncament per sortides petites).<\/li>\n\n\n\n
                  • Si \u00e9s 0 bytes<\/strong> \u2192 Piper mateix descarta els textos curts.<\/li>\n<\/ul>\n\n\n\n

                    Diagn\u00f2stic 3: Troba el llindar<\/h2>\n\n\n\n

                    bash<\/p>\n\n\n\n

                    1for T in \u00abteulada\u00bb \u00abHola\u00bb \u00abHola Ona\u00bb \u00abHola Ona, com estas\u00bb \u00abAquesta es una frase normal\u00bb \u00abAquesta es una frase una mica mes llarga per provar el sistema\u00bb; do<\/p>\n\n\n\n

                    2 printf \u00ab%-45s \u2192 \u00bb \u00ab$T\u00bb<\/p>\n\n\n\n

                    3 docker exec piper bash -c \u00abecho \\\u00bb$T\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | wc -c<\/p>\n\n\n\n

                    4done<\/p>\n\n\n\n

                    5<\/p>\n\n\n\n

                    5 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                    Aix\u00f2 et dir\u00e0 a partir de quants car\u00e0cters comen\u00e7a a generar \u00e0udio (si \u00e9s que ho fa).<\/p>\n\n\n\n

                    Diagn\u00f2stic 4: Mira si piper escriu algun error a stderr<\/h2>\n\n\n\n

                    bash<\/p>\n\n\n\n

                    1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb 2>&1 | head -50<\/p>\n\n\n\n

                    2<\/p>\n\n\n\n

                    2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                    \n\n\n\n

                    Executa el diagn\u00f2stic 1 i 2<\/strong> i digue’m els resultats \u2014 amb aix\u00f2 ja sabrem si el culpable \u00e9s Piper o el canal de sortida de docker exec. Segons el resultat, la soluci\u00f3 ser\u00e0 ben diferent:<\/p>\n\n\n\n

                      \n
                    • Piper no genera res<\/strong> \u2192 caldr\u00e0 ajustar com es passa el text (potser la versi\u00f3 de piper del contenidor t\u00e9 un m\u00ednim de car\u00e0cters, o requereix un par\u00e0metre com --sentence-silence<\/code> o for\u00e7ar l’\u00faltima frase).<\/li>\n\n\n\n
                    • docker exec perd la sortida<\/strong> \u2192 la soluci\u00f3 ser\u00e0 escriure a un fitxer dins el contenidor i fer docker cp<\/code>, o afegir sleep<\/code> al final perqu\u00e8 docker exec flushegi.<\/li>\n<\/ul>\n\n\n\n

                      Enrutamiento Inteligente<\/p>\n\n\n\n

                      Miquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | wc -c 0 Miquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/test.wav && wc -c \/tmp\/test.wav\u00bb 28204 \/tmp\/test.wav Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                      Deep think completado (tard\u00f3 50s)<\/p>\n\n\n\n

                      Resultat molt important! Aix\u00f2 canvia completament el diagn\u00f2stic:<\/p>\n\n\n\n

                      Prova<\/th>Resultat<\/th><\/tr><\/thead>
                      --output_file -<\/code> (stdout)<\/td>0 bytes<\/strong><\/td><\/tr>
                      --output_file \/tmp\/test.wav<\/code> (fitxer)<\/td>28.204 bytes<\/strong> \u2705<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                      Piper S\u00cd que sintetitza \u00abteulada.\u00bb perfectament<\/strong> (28 KB \u2248 0,6 s d’\u00e0udio). El problema no \u00e9s la s\u00edntesi: \u00e9s que el canal de sortida per stdout (--output_file -<\/code>) a trav\u00e9s de docker exec<\/code> es perd<\/strong>.<\/p>\n\n\n\n

                      Qu\u00e8 passa exactament?<\/h2>\n\n\n\n

                      Quan el text \u00e9s curt, la s\u00edntesi \u00e9s tan r\u00e0pida que el contenidor acaba abans que docker exec hagi flushejat\/entregat la sortida de stdout<\/strong>, i el flux es tanca amb 0 bytes. Amb el text llarg, hi ha prou temps\/dades perqu\u00e8 la sortida arribi b\u00e9 \u2014 per aix\u00f2 el primer comandament va funcionar.<\/p>\n\n\n\n

                      No \u00e9s culpa de Piper (sintetitza b\u00e9), sin\u00f3 de la combinaci\u00f3 stdout + docker exec<\/code> per a processos que acaben molt de pressa<\/strong>.<\/p>\n\n\n\n

                      Comprovacions per confirmar-ho<\/h2>\n\n\n\n

                      1. El text llarg encara surt per stdout?<\/strong><\/p>\n\n\n\n

                      bash<\/p>\n\n\n\n

                      1docker exec piper bash -c \u00abecho \\\u00bbHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | wc -c<\/p>\n\n\n\n

                      2<\/p>\n\n\n\n

                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                      Si d\u00f3na >0, confirma que \u00e9s una q\u00fcesti\u00f3 de temps\/mida de la sortida.<\/p>\n\n\n\n

                      2. Hi ha un fitxer anomenat -<\/code> dins el contenidor?<\/strong><\/p>\n\n\n\n

                      bash<\/p>\n\n\n\n

                      1docker exec piper bash -c \u00abls -la\u00bb<\/p>\n\n\n\n

                      2<\/p>\n\n\n\n

                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                      (Per descartar que piper interpreti -<\/code> com a nom de fitxer.)<\/p>\n\n\n\n

                      Soluci\u00f3 robusta (recomanada)<\/h2>\n\n\n\n

                      Ja que sabem que escriure a fitxer funciona sempre, la soluci\u00f3 fiable \u00e9s generar el WAV dins del contenidor i copiar-lo fora<\/strong>:<\/p>\n\n\n\n

                      bash<\/p>\n\n\n\n

                      1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/veu.wav\u00bb<\/p>\n\n\n\n

                      2docker cp piper:\/tmp\/veu.wav \/tmp\/veu.wav<\/p>\n\n\n\n

                      3aplay \/tmp\/veu.wav<\/p>\n\n\n\n

                      4<\/p>\n\n\n\n

                      4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                      I per al teu servidor TTS, un script com aquest:<\/p>\n\n\n\n

                      bash<\/p>\n\n\n\n

                      1#!\/bin\/bash<\/p>\n\n\n\n

                      2# tts.sh \u00abtext a dir\u00bb<\/p>\n\n\n\n

                      3TEXT=\u00bb${1:?\u00das: tts.sh ‘text’}\u00bb<\/p>\n\n\n\n

                      4[[ \u00ab$TEXT\u00bb =~ [.!?]$ ]] || TEXT=\u00bb$TEXT.\u00bb<\/p>\n\n\n\n

                      5<\/p>\n\n\n\n

                      6docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/veu.wav\u00bb<\/p>\n\n\n\n

                      7docker cp piper:\/tmp\/veu.wav \/tmp\/veu.wav<\/p>\n\n\n\n

                      8aplay \/tmp\/veu.wav<\/p>\n\n\n\n

                      9<\/p>\n\n\n\n

                      9 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                      Alternativa si vols mantenir el pipe<\/h2>\n\n\n\n

                      Espera un moment despr\u00e9s de la s\u00edntesi dins el contenidor perqu\u00e8 docker exec flushegi la sortida:<\/p>\n\n\n\n

                      bash<\/p>\n\n\n\n

                      1docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -; sleep 1\u00bb | aplay -f S16_LE -r 22050 -c 1<\/p>\n\n\n\n

                      2<\/p>\n\n\n\n

                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                      El sleep 1<\/code> fa que el contenidor no surti immediatament i la sortida de stdout arribi sencera.<\/p>\n\n\n\n

                      \n\n\n\n

                      Prova primer la comprovaci\u00f3 1<\/strong> (text llarg per wc) per confirmar la teoria, i despr\u00e9s la soluci\u00f3 amb docker cp<\/code> que \u00e9s la que et resoldr\u00e0 el problema de manera definitiva. Qu\u00e8 surt?<\/p>\n\n\n\n

                      Enrutamiento Inteligente<\/p>\n\n\n\n

                      Miquel@Miquel-PC:~$ aplay \/tmp\/veu.wav Playing WAVE ‘\/tmp\/veu.wav’ : Signed 16 bit Little Endian, Rate 22050 Hz, Mono Miquel@Miquel-PC:~$ #!\/bin\/bash # tts.sh \u00abtext a dir\u00bb TEXT=\u00bb${1:?\u00das: tts.sh ‘text’}\u00bb [[ \u00ab$TEXT\u00bb =~ [.!?]$ ]] || TEXT=\u00bb$TEXT.\u00bb docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/veu.wav\u00bb docker cp piper:\/tmp\/veu.wav \/tmp\/veu.wav aplay \/tmp\/veu.wav bash: 1: \u00das: tts.sh text bash: !?]$ ]] || TEXT=\u00bb$TEXT.\u00bb: event not found Traceback (most recent call last): File \u00ab\/lsiopy\/bin\/piper\u00bb, line 6, in <module> sys.exit(main()) ~~~~^^ File \u00ab\/lsiopy\/lib\/python3.14\/site-packages\/piper\/__main__.py\u00bb, line 251, in main with wav_file: ^^^^^^^^ File \u00ab\/usr\/lib\/python3.14\/wave.py\u00bb, line 477, in __exit__ self.close() ~~~~~~~~~~^^ File \u00ab\/usr\/lib\/python3.14\/wave.py\u00bb, line 595, in close self._ensure_header_written(0) ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^ File \u00ab\/usr\/lib\/python3.14\/wave.py\u00bb, line 613, in _ensure_header_written raise Error(‘# channels not specified’) wave.Error: # channels not specified Successfully copied 1.54kB to \/tmp\/veu.wav Playing raw data ‘\/tmp\/veu.wav’ : Unsigned 8 bit, Rate 8000 Hz, Mono Miquel@Miquel-PC:~$ el primer bash ha dit ‘teulada’ el segon nom\u00e9s un sorollet<\/p>\n\n\n\n

                      Deep think completado (tard\u00f3 15s)<\/p>\n\n\n\n

                      Perfecte! El que has fet \u00e9s enganxar el script l\u00ednia a l\u00ednia al terminal interactiu<\/strong>, i aix\u00f2 ha provocat dos errors cl\u00e0ssics:<\/p>\n\n\n\n

                      Qu\u00e8 ha passat<\/h2>\n\n\n\n
                        \n
                      1. bash: !?]$ ]]...: event not found<\/code><\/strong> \u2192 El terminal interactiu de bash interpreta !?<\/code> com a expansi\u00f3 d’historial<\/em> (buscar l’\u00faltima comanda que contingui aquest text). Per aix\u00f2 la l\u00ednia del regex ha esclatat. En un fitxer de script aix\u00f2 no passa<\/strong> (l’expansi\u00f3 d’historial nom\u00e9s actua al terminal).<\/li>\n\n\n\n
                      2. wave.Error: # channels not specified<\/code><\/strong> \u2192 Com que l’assignaci\u00f3 de $TEXT<\/code> va fallar (no hi havia argument), Piper va rebre text buit<\/strong> i va petar en intentar tancar el WAV. El docker cp<\/code> va copiar un fitxer antic\/parcial (1,54 kB) i aplay el va reproduir com a dades brutes \u2192 el \u00absorollet\u00bb.<\/li>\n<\/ol>\n\n\n\n

                        La soluci\u00f3: crea el script com a fitxer<\/h2>\n\n\n\n

                        Copia i enganxa aix\u00f2 (sencer) al terminal:<\/p>\n\n\n\n

                        bash<\/p>\n\n\n\n

                        1cat > ~\/tts.sh <<‘EOF’<\/p>\n\n\n\n

                        2#!\/bin\/bash<\/p>\n\n\n\n

                        3# tts.sh \u00abtext a dir\u00bb<\/p>\n\n\n\n

                        4TEXT=\u00bb$1″<\/p>\n\n\n\n

                        5if [ -z \u00ab$TEXT\u00bb ]; then<\/p>\n\n\n\n

                        6 echo \u00ab\u00das: tts.sh ‘text a dir'\u00bb >&2<\/p>\n\n\n\n

                        7 exit 1<\/p>\n\n\n\n

                        8fi<\/p>\n\n\n\n

                        9<\/p>\n\n\n\n

                        10# Afegeix un punt final si el text no en t\u00e9<\/p>\n\n\n\n

                        11case \u00ab$TEXT\u00bb in<\/p>\n\n\n\n

                        12 *[.!?]) ;;<\/p>\n\n\n\n

                        13 *) TEXT=\u00bb$TEXT.\u00bb ;;<\/p>\n\n\n\n

                        14esac<\/p>\n\n\n\n

                        15<\/p>\n\n\n\n

                        16docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/veu.wav\u00bb<\/p>\n\n\n\n

                        17docker cp piper:\/tmp\/veu.wav \/tmp\/veu.wav<\/p>\n\n\n\n

                        18aplay \/tmp\/veu.wav<\/p>\n\n\n\n

                        19EOF<\/p>\n\n\n\n

                        20chmod +x ~\/tts.sh<\/p>\n\n\n\n

                        21<\/p>\n\n\n\n

                        21 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                        Ara prova-ho:<\/p>\n\n\n\n

                        bash<\/p>\n\n\n\n

                        1~\/tts.sh \u00abteulada\u00bb<\/p>\n\n\n\n

                        2~\/tts.sh \u00abHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\u00bb<\/p>\n\n\n\n

                        3<\/p>\n\n\n\n

                        3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                        Per qu\u00e8 aquesta versi\u00f3 funciona<\/h2>\n\n\n\n
                          \n
                        • case<\/code> en comptes de regex<\/strong> \u2192 evita el problema del !<\/code> (l’expansi\u00f3 d’historial no s’activa dins d’un fitxer).<\/li>\n\n\n\n
                        • El text es passa com a argument<\/strong> ($1<\/code>), no enganxat al mig.<\/li>\n\n\n\n
                        • La s\u00edntesi va a un fitxer<\/strong> (\/tmp\/veu.wav<\/code>), que \u00e9s el m\u00e8tode que hem verificat que sempre funciona (28 KB per a \u00abteulada.\u00bb), i despr\u00e9s docker cp<\/code> el porta a l’host perqu\u00e8 aplay<\/code> el reprodueixi.<\/li>\n<\/ul>\n\n\n\n

                          Si vols confirmar la teoria del stdout perdut, tamb\u00e9 pots provar aix\u00f2 (hauria de donar >0):<\/p>\n\n\n\n

                          bash<\/p>\n\n\n\n

                          1docker exec piper bash -c \u00abecho \\\u00bbHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | wc -c<\/p>\n\n\n\n

                          2<\/p>\n\n\n\n

                          2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                          Prova el script i digue’m si ja parla l’Ona correctament! \ud83d\udde3\ufe0f<\/p>\n\n\n\n

                          Enrutamiento Inteligente<\/p>\n\n\n\n

                          Miquel@Miquel-PC:~$ ~\/tts.sh \u00abteulada\u00bb ~\/tts.sh \u00abHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\u00bb Successfully copied 30.2kB to \/tmp\/veu.wav Playing WAVE ‘\/tmp\/veu.wav’ : Signed 16 bit Little Endian, Rate 22050 Hz, Mono Successfully copied 214kB to \/tmp\/veu.wav Playing WAVE ‘\/tmp\/veu.wav’ : Signed 16 bit Little Endian, Rate 22050 Hz, Mono Miquel@Miquel-PC:~$ perfecte. els dos van b\u00e9<\/p>\n\n\n\n

                          Deep think completado (tard\u00f3 2s)<\/p>\n\n\n\n

                          Molt b\u00e9! \ud83c\udf89 L’Ona ja parla sense problemes, tant per a paraules curtes (teulada<\/code> \u2192 30,2 kB) com per a frases llargues (214 kB).<\/p>\n\n\n\n

                          Resum del problema i la soluci\u00f3<\/h2>\n\n\n\n

                          Causa arrel:<\/strong> el canal de sortida per stdout (--output_file -<\/code>) a trav\u00e9s de docker exec<\/code> perdia les dades quan la s\u00edntesi acabava massa de pressa (textos curts). No era culpa ni de Piper ni de aplay.<\/p>\n\n\n\n

                          Soluci\u00f3 final:<\/strong> generar el WAV dins del contenidor, copiar-lo amb docker cp<\/code> i reproduir-lo amb aplay<\/code>. Aix\u00ed sempre funciona, sigui quin sigui el llarg del text.<\/p>\n\n\n\n

                          El teu script ~\/tts.sh<\/code> queda com a base s\u00f2lida per al teu servidor TTS.<\/p>\n\n\n\n

                          Millores opcionals (si vols)<\/h2>\n\n\n\n
                            \n
                          1. Evitar el docker cp<\/code><\/strong> \u2014 Si recrees el contenidor, pots muntar un volum compartit i escriure-hi directament:<\/li>\n<\/ol>\n\n\n\n

                            bash<\/p>\n\n\n\n

                            1docker run … -v \/tmp\/audio:\/audio … piper<\/p>\n\n\n\n

                            2# despr\u00e9s, al script:<\/p>\n\n\n\n

                            3docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model … –output_file \/audio\/veu.wav\u00bb<\/p>\n\n\n\n

                            4aplay \/tmp\/audio\/veu.wav<\/p>\n\n\n\n

                            5<\/p>\n\n\n\n

                            5 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                              \n
                            1. Tornar al pipe pur<\/strong> (si vols estalviar el pas per fitxer), amb el sleep<\/code> perqu\u00e8 docker exec flushegi:<\/li>\n<\/ol>\n\n\n\n

                              bash<\/p>\n\n\n\n

                              1docker exec piper bash -c \u00abecho \\\u00bb$TEXT\\\u00bb | piper –model … –output_file -; sleep 1″ | aplay -f S16_LE -r 22050 -c 1<\/p>\n\n\n\n

                              2<\/p>\n\n\n\n

                              2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                \n
                              1. Integrar-ho amb el teu tts-server<\/strong>: el script ja accepta el text com a argument, aix\u00ed que pots cridar-lo des del teu servei quan rebi una petici\u00f3.<\/li>\n<\/ol>\n\n\n\n

                                Necessites res m\u00e9s, com ara adaptar el script perqu\u00e8 llegeixi el text d’un fitxer o d’un par\u00e0metre de la l\u00ednia de comandes d’una altra manera?<\/p>\n\n\n\n

                                Enrutamiento Inteligente<\/p>\n\n\n\n

                                dacord ho integrem a tts_server<\/p>\n\n\n\n

                                Deep think completado (tard\u00f3 2s)<\/p>\n\n\n\n

                                D’acord! Abans de proposar-te la integraci\u00f3, deixa’m recordar si tenim context sobre el teu tts-server.<\/p>\n\n\n\n

                                Deep think completado (tard\u00f3 2s)<\/p>\n\n\n\n

                                No tinc cap informaci\u00f3 pr\u00e8via sobre el teu tts-server<\/code>, aix\u00ed que necessito que em diguis com est\u00e0 fet per integrar-hi el script de la manera correcta.<\/p>\n\n\n\n

                                Preguntes r\u00e0pides<\/h2>\n\n\n\n
                                  \n
                                1. Qu\u00e8 \u00e9s tts-server<\/code>?<\/strong> \u2014 Un script Python (Flask\/FastAPI)? Node.js? Un binari? Un servei systemd?<\/li>\n\n\n\n
                                2. On \u00e9s el codi?<\/strong> \u2014 Ruta del fitxer o directori (p. ex. ~\/tts-server\/<\/code>, \/opt\/tts-server\/<\/code>…)<\/li>\n\n\n\n
                                3. Quina interf\u00edcie exposa?<\/strong> \u2014 HTTP (endpoint tipus POST \/tts<\/code>)? L\u00ednia de comandes? Socket?<\/li>\n\n\n\n
                                4. Com rep el text i qu\u00e8 ha de retornar?<\/strong> \u2014 JSON amb text? Ha de tornar l’\u00e0udio com a resposta HTTP o nom\u00e9s reproduir-lo?<\/li>\n<\/ol>\n\n\n\n

                                  Mentrestant, una integraci\u00f3 gen\u00e8rica (Python + HTTP)<\/h2>\n\n\n\n

                                  Perqu\u00e8 et facis una idea, aix\u00ed seria un servidor m\u00ednim que fa servir el teu pipeline (script \u2192 Piper \u2192 aplay) i retorna l’\u00e0udio:<\/p>\n\n\n\n

                                  python<\/p>\n\n\n\n

                                  1from flask import Flask, request, Response, jsonify<\/p>\n\n\n\n

                                  2import subprocess, tempfile, os<\/p>\n\n\n\n

                                  3<\/p>\n\n\n\n

                                  4app = Flask(__name__)<\/p>\n\n\n\n

                                  5<\/p>\n\n\n\n

                                  6def sintetitza(text):<\/p>\n\n\n\n

                                  7 \u00ab\u00bb\u00bbGenera WAV amb el contenidor piper i el retorna com a bytes.\u00bb\u00bb\u00bb<\/p>\n\n\n\n

                                  8 # 1. Genera el WAV dins del contenidor<\/p>\n\n\n\n

                                  9 subprocess.run(<\/p>\n\n\n\n

                                  10 [\u00abdocker\u00bb, \u00abexec\u00bb, \u00abpiper\u00bb, \u00abbash\u00bb, \u00ab-c\u00bb,<\/p>\n\n\n\n

                                  11 f\u00bbecho \\\u00bb{text}\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/veu.wav\u00bb],<\/p>\n\n\n\n

                                  12 check=True)<\/p>\n\n\n\n

                                  13 # 2. Copia’l fora<\/p>\n\n\n\n

                                  14 subprocess.run([\u00abdocker\u00bb, \u00abcp\u00bb, \u00abpiper:\/tmp\/veu.wav\u00bb, \u00ab\/tmp\/veu.wav\u00bb], check=True)<\/p>\n\n\n\n

                                  15 # 3. Llegeix el fitxer<\/p>\n\n\n\n

                                  16 with open(\u00ab\/tmp\/veu.wav\u00bb, \u00abrb\u00bb) as f:<\/p>\n\n\n\n

                                  17 return f.read()<\/p>\n\n\n\n

                                  18<\/p>\n\n\n\n

                                  19@app.post(\u00ab\/tts\u00bb)<\/p>\n\n\n\n

                                  20def tts():<\/p>\n\n\n\n

                                  21 data = request.get_json(force=True)<\/p>\n\n\n\n

                                  22 text = data.get(\u00abtext\u00bb, \u00ab\u00bb)<\/p>\n\n\n\n

                                  23 if not text:<\/p>\n\n\n\n

                                  24 return jsonify({\u00aberror\u00bb: \u00abfalta el camp ‘text'\u00bb}), 400<\/p>\n\n\n\n

                                  25 audio = sintetitza(text)<\/p>\n\n\n\n

                                  26 return Response(audio, mimetype=\u00bbaudio\/wav\u00bb)<\/p>\n\n\n\n

                                  27<\/p>\n\n\n\n

                                  27 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                  Amb aix\u00f2, una petici\u00f3 POST \/tts<\/code> amb {\"text\": \"teulada\"}<\/code> retornaria el WAV directament, i tu decideixes si el client el reprodueix o el servidor ho fa amb aplay<\/code>.<\/p>\n\n\n\n

                                  Expliqueu-me com \u00e9s el teu tts-server<\/code> i t’ho adapto exactament. \ud83d\ude42<\/p>\n\n\n\n

                                  Enrutamiento Inteligente<\/p>\n\n\n\n

                                  \"\"\/<\/figure>\n\n\n\n

                                  demana’m la informaci\u00f3 que vulguis<\/p>\n\n\n\n

                                  Deep think completado (tard\u00f3 5s)<\/p>\n\n\n\n

                                  Gr\u00e0cies per la captura! Ja veig que \u00e9s un projecte Python amb venv<\/code>, uns scripts de prova (prova_piper.py<\/code>, prova_piper_dbg.py<\/code>) i una subcarpeta tts-server\/<\/code>. Per integrar-hi el pipeline correctament, necessito que em responguis aix\u00f2:<\/p>\n\n\n\n

                                  Preguntes<\/h2>\n\n\n\n

                                  1. Ruta absoluta del projecte<\/strong>
                                  Quina \u00e9s la ruta completa del directori? (p. ex. \/home\/Miquel\/tts-server\/<\/code>) \u2014 Aix\u00ed puc llegir els fitxers directament i veure el codi real.<\/p>\n\n\n\n

                                  2. On \u00e9s el codi principal del servidor?<\/strong><\/p>\n\n\n\n

                                    \n
                                  • El servidor en si \u00e9s dins la subcarpeta tts-server\/<\/code>? (quins fitxers hi ha?)<\/li>\n\n\n\n
                                  • O prova_piper.py<\/code> \/ prova_piper_dbg.py<\/code> s\u00f3n els scripts que fas servir?<\/li>\n<\/ul>\n\n\n\n

                                    3. Quin framework utilitza?<\/strong>
                                    Flask? FastAPI? http.server<\/code> de Python? Sockets?<\/p>\n\n\n\n

                                    4. Quina interf\u00edcie exposa?<\/strong><\/p>\n\n\n\n

                                      \n
                                    • Port i endpoints (p. ex. POST \/tts<\/code> al port 5000)<\/li>\n\n\n\n
                                    • Qu\u00e8 rep (JSON amb text<\/code>?) i qu\u00e8 retorna (\u00e0udio WAV? nom\u00e9s OK? text?)<\/li>\n<\/ul>\n\n\n\n

                                      5. Com l’arrenques?<\/strong>
                                      python server.py<\/code>? Un servei systemd? Docker?<\/p>\n\n\n\n

                                      6. El venv\/<\/code> \u00e9s el que utilitza el servidor?<\/strong>
                                      I quin Python hi ha dins (venv\/bin\/python --version<\/code>)?<\/p>\n\n\n\n

                                      \n\n\n\n

                                      Un cop em diguis la ruta absoluta<\/strong>, puc llegir prova_piper.py<\/code>, server.log<\/code> i el contingut de tts-server\/<\/code> per adaptar-ho exactament al teu codi, sense suposar res.<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      ubicaci\u00f3: \/home\/Miquel\/tts-server<\/p>\n\n\n\n

                                      Execution ProcessProcessed for 24s<\/p>\n\n\n\n

                                      Perfecte, ara llegeixo els fitxers per veure el codi real.<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 8s)<\/p>\n\n\n\n

                                      Ja tinc una visi\u00f3 clara! Llegeixo els fitxers de prova i el log: el teu tts-server<\/code> \u00e9s en realitat un servidor speech-to-speech<\/strong> (STT amb Parakeet + VAD + TTS amb Piper) que escolta al port 10200<\/strong> i parla per socket amb missatges JSON (synthesize<\/code>, audio-start<\/code>, audio-chunk<\/code>, audio-stop<\/code>).<\/p>\n\n\n\n

                                      Els fitxers server.py<\/code>\/app.py<\/code> dins la subcarpeta no existeixen (m’ha dit \u00abPar\u00e0metre no v\u00e0lid\u00bb). Deixa’m localitzar el codi real del servidor.<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 3s)<\/p>\n\n\n\n

                                      La subcarpeta tts-server\/<\/code> i piper-config\/<\/code> existeixen, per\u00f2 no puc llistar-ne el contingut directament des d’aqu\u00ed (nom\u00e9s en veig les metadades). Necessito que em mostris l’arbre de fitxers per localitzar el codi del servidor.<\/p>\n\n\n\n

                                      Executa aix\u00f2 al terminal i enganxa’m la sortida<\/h2>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1find \/home\/Miquel\/tts-server -maxdepth 4 -type f | sort<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      Amb la llista de fitxers podr\u00e9 llegir el codi exacte on es crida Piper dins del servidor (el que ara deu fer servir --output_file -<\/code> i pateix el problema dels textos curts) i adaptar-hi la soluci\u00f3 robusta del fitxer + docker cp<\/code>.<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      Miquel@Miquel-PC:~$ find \/home\/Miquel\/tts-server -maxdepth 4 -type f | sort \/home\/Miquel\/tts-server\/piper-config\/ca_ES-upc_ona-medium.onnx \/home\/Miquel\/tts-server\/piper-config\/ca_ES-upc_ona-medium.onnx.json \/home\/Miquel\/tts-server\/piper-config\/voices.json \/home\/Miquel\/tts-server\/prova_piper_dbg.py \/home\/Miquel\/tts-server\/prova_piper.py \/home\/Miquel\/tts-server\/server.log \/home\/Miquel\/tts-server\/tts-server\/prova_piper_dbg.py \/home\/Miquel\/tts-server\/tts-server\/prova_piper.py \/home\/Miquel\/tts-server\/tts-server\/server.log \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/accelerate \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/accelerate-config \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/accelerate-estimate-memory \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/accelerate-launch \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/accelerate-merge-weights \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/activate \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/activate.csh \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/activate.fish \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/Activate.ps1 \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/cffi-gen-src \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/distro \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/f2py \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/fastapi \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/faster-qwen3-tts \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/gradio \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/hf \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/hf-gradio \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/httpx \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/huggingface-cli \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/idna \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/isympy \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/markdown-it \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/nano-parakeet \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/nltk \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/normalizer \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/numba \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/numpy-config \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/onnxruntime_test \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/openai \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/pip \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/pip3 \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/pip3.12 \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/proton \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/proton-viewer \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/pygmentize \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/qwen-tts-demo \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/speech-to-speech \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/tiny-agents \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/torchfrtrace \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/torchrun \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/tqdm \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/transformers \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/typer \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/upload_theme \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/uvicorn \/home\/Miquel\/tts-server\/tts-server\/venv\/bin\/websockets \/home\/Miquel\/tts-server\/tts-server\/venv\/pyvenv.cfg \/home\/Miquel\/tts-server\/venv\/bin\/accelerate \/home\/Miquel\/tts-server\/venv\/bin\/accelerate-config \/home\/Miquel\/tts-server\/venv\/bin\/accelerate-estimate-memory \/home\/Miquel\/tts-server\/venv\/bin\/accelerate-launch \/home\/Miquel\/tts-server\/venv\/bin\/accelerate-merge-weights \/home\/Miquel\/tts-server\/venv\/bin\/activate \/home\/Miquel\/tts-server\/venv\/bin\/activate.csh \/home\/Miquel\/tts-server\/venv\/bin\/activate.fish \/home\/Miquel\/tts-server\/venv\/bin\/Activate.ps1 \/home\/Miquel\/tts-server\/venv\/bin\/cffi-gen-src \/home\/Miquel\/tts-server\/venv\/bin\/distro \/home\/Miquel\/tts-server\/venv\/bin\/f2py \/home\/Miquel\/tts-server\/venv\/bin\/fastapi \/home\/Miquel\/tts-server\/venv\/bin\/faster-qwen3-tts \/home\/Miquel\/tts-server\/venv\/bin\/gradio \/home\/Miquel\/tts-server\/venv\/bin\/hf \/home\/Miquel\/tts-server\/venv\/bin\/hf-gradio \/home\/Miquel\/tts-server\/venv\/bin\/httpx \/home\/Miquel\/tts-server\/venv\/bin\/huggingface-cli \/home\/Miquel\/tts-server\/venv\/bin\/idna \/home\/Miquel\/tts-server\/venv\/bin\/isympy \/home\/Miquel\/tts-server\/venv\/bin\/markdown-it \/home\/Miquel\/tts-server\/venv\/bin\/nano-parakeet \/home\/Miquel\/tts-server\/venv\/bin\/nltk \/home\/Miquel\/tts-server\/venv\/bin\/normalizer \/home\/Miquel\/tts-server\/venv\/bin\/numba \/home\/Miquel\/tts-server\/venv\/bin\/numpy-config \/home\/Miquel\/tts-server\/venv\/bin\/onnxruntime_test \/home\/Miquel\/tts-server\/venv\/bin\/openai \/home\/Miquel\/tts-server\/venv\/bin\/pip \/home\/Miquel\/tts-server\/venv\/bin\/pip3 \/home\/Miquel\/tts-server\/venv\/bin\/pip3.12 \/home\/Miquel\/tts-server\/venv\/bin\/proton \/home\/Miquel\/tts-server\/venv\/bin\/proton-viewer \/home\/Miquel\/tts-server\/venv\/bin\/pygmentize \/home\/Miquel\/tts-server\/venv\/bin\/qwen-tts-demo \/home\/Miquel\/tts-server\/venv\/bin\/speech-to-speech \/home\/Miquel\/tts-server\/venv\/bin\/tiny-agents \/home\/Miquel\/tts-server\/venv\/bin\/torchfrtrace \/home\/Miquel\/tts-server\/venv\/bin\/torchrun \/home\/Miquel\/tts-server\/venv\/bin\/tqdm \/home\/Miquel\/tts-server\/venv\/bin\/transformers \/home\/Miquel\/tts-server\/venv\/bin\/typer \/home\/Miquel\/tts-server\/venv\/bin\/upload_theme \/home\/Miquel\/tts-server\/venv\/bin\/uvicorn \/home\/Miquel\/tts-server\/venv\/bin\/websockets \/home\/Miquel\/tts-server\/venv\/pyvenv.cfg Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 6s)<\/p>\n\n\n\n

                                      El codi del servidor no \u00e9s un projecte local<\/strong>: \u00e9s un paquet Python instal\u00b7lat al venv (el binari speech-to-speech<\/code> i el log ho confirmen \u2014 paquet speech_to_speech<\/code>). El find<\/code> amb -maxdepth 4<\/code> no arriba fins a site-packages<\/code>, per aix\u00f2 no el veus.<\/p>\n\n\n\n

                                      Executa aix\u00f2 per localitzar el codi on es crida Piper:<\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1find \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech -type f -name \u00ab*.py\u00bb | sort<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1grep -rln \u00abpiper\\|piper-config\\|output_file\u00bb \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/ 2>\/dev\/null<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      La primera comanda et donar\u00e0 l’estructura del paquet; la segona, quins fitxers concrets fan servir Piper<\/strong> (all\u00e0 hi haur\u00e0 el codi que hem de modificar per arreglar els textos curts). Enganxa’m la sortida.<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      Miquel@Miquel-PC:~$ find \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech -type f -name \u00ab*.py\u00bb | sort \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/audio.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/base.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/conversation.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/response.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/handlers\/session.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/pipeline_unit.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/runtime_config.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/server.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/service.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/utils.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/api\/openai_realtime\/websocket_router.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/chat_tts_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/facebookmms_tts_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/faster_whisper_stt_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/kokoro_tts_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/language_model_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/language_model_base_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/mlx_audio_whisper_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/module_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/paraformer_stt_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/parakeet_tdt_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/pocket_tts_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/qwen3_tts_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/responses_api_language_model_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/socket_receiver_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/socket_sender_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/vad_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/websocket_streamer_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/arguments_classes\/whisper_stt_arguments.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/baseHandler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/connections\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/connections\/local_audio_streamer.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/connections\/socket_receiver.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/connections\/socket_sender.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/connections\/websocket_streamer.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/chat.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/compaction_prompt.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/language_model.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/lm_output_processor.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/responses_api_language_model.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/tool_call\/function_call.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/tool_call\/function_tool.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/tool_call\/signature_from_schema.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/tool_call\/tool_prompt.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/utils.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/LLM\/voice_prompt.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/cancel_scope.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/control.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/events.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/handler_types.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/log_context.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/messages.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/queue_types.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/pipeline\/speculative_turns.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/s2s_pipeline.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/base_stt_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/faster_whisper_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/lightning_whisper_mlx_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/mlx_audio_whisper_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/paraformer_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/parakeet_tdt_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/smart_progressive_streaming.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/transcription_notifier.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/STT\/whisper_stt_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/chatTTS_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/facebookmms_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/kokoro_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/pocket_tts_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/TTS\/qwen3_tts_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/utils\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/utils\/mlx_lock.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/utils\/thread_manager.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/utils\/utils.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/VAD\/__init__.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/VAD\/vad_handler.py \/home\/Miquel\/tts-server\/venv\/lib\/python3.12\/site-packages\/speech_to_speech\/VAD\/vad_iterator.py Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 10s)<\/p>\n\n\n\n

                                      Interessant! Al paquet speech_to_speech<\/code> no hi ha cap handler de Piper<\/strong> (nom\u00e9s chatTTS, facebookmms, kokoro, pocket_tts, qwen3). Per\u00f2 el teu client prova_piper.py<\/code> fa servir la veu ca_ES-upc_ona-medium<\/code> (Piper) i funciona. Deixa’m llegir els fitxers clau per veure com es connecta Piper i on es genera l’\u00e0udio.<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 8s)<\/p>\n\n\n\n

                                      El paquet speech_to_speech<\/code> instal\u00b7lat \u00e9s el pipeline complet (VAD \u2192 STT \u2192 LLM \u2192 TTS amb chatTTS\/facebookMMS\/pocket\/kokoro\/qwen3), per\u00f2 no cont\u00e9 cap handler de Piper<\/strong> ni el protocol JSON que fa servir el teu prova_piper.py<\/code> (missatges synthesize<\/code>, audio-chunk<\/code>, port 10200).<\/p>\n\n\n\n

                                      Aix\u00f2 vol dir que el teu servidor a 10200 \u00e9s un codi propi<\/strong> que no he trobat encara. Necessito localitzar-lo.<\/p>\n\n\n\n

                                      Executa aquestes comandes i enganxa’m la sortida<\/h2>\n\n\n\n

                                      1. Trobar el codi que implementa el protocol (synthesize \/ audio-chunk \/ 10200):<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1grep -rln \u00abaudio-chunk\\|synthesize\\|10200\u00bb \/home\/Miquel\/tts-server 2>\/dev\/null | grep -v \u00ab\/venv\/\u00bb<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      Si no surt res, amplia a tot el teu home:<\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1grep -rln \u00abaudio-chunk\u00bb \/home\/Miquel 2>\/dev\/null | grep -v \u00ab\/venv\/\u00bb | head -20<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      2. Veure qu\u00e8 est\u00e0 executant-se ara mateix (el servidor en marxa):<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1ps aux | grep -v grep | grep -E \u00abpython|speech\u00bb<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      3. Com l’arrenques? (historial de comandes):<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1history | grep -iE \u00abtts|speech|10200|piper\u00bb | tail -20<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      Amb aix\u00f2 sabr\u00e9 on \u00e9s el codi del servidor que hem de modificar (on es crida Piper amb --output_file -<\/code>) per canviar-lo a la soluci\u00f3 robusta del fitxer + docker cp<\/code>.<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      Miquel@Miquel-PC:~$ grep -rln \u00abaudio-chunk\\|synthesize\\|10200\u00bb \/home\/Miquel\/tts-server 2>\/dev\/null | grep -v \u00ab\/venv\/\u00bb<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 2s)<\/p>\n\n\n\n

                                      La cerca dins de \/home\/Miquel\/tts-server<\/code> no ha trobat res, aix\u00ed que el codi del servidor de 10200 \u00e9s en un altre lloc<\/strong>. Necessito les altres dues comandes per localitzar-lo:<\/p>\n\n\n\n

                                      1. Cerca m\u00e9s \u00e0mplia a tot el teu home:<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1grep -rln \u00abaudio-chunk\u00bb \/home\/Miquel 2>\/dev\/null | grep -v \u00ab\/venv\/\u00bb | head -20<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      2. Qu\u00e8 s’est\u00e0 executant ara mateix?<\/strong> (aix\u00f2 \u00e9s el m\u00e9s important \u2014 et dir\u00e0 des d’on s’ha llan\u00e7at el servidor):<\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1ps aux | grep -v grep | grep -E \u00abpython|speech\u00bb<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      3. Historial d’arrencada:<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1history | grep -iE \u00abtts|speech|10200|piper\u00bb | tail -20<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      La comanda 2<\/strong> \u00e9s clau: mostrar\u00e0 la l\u00ednia de comandes completa del proc\u00e9s que escolta al port 10200, i amb aix\u00f2 sabr\u00e9 exactament quin script i des de quina ruta s’executa.<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      tampoc trova res.<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 3s)<\/p>\n\n\n\n

                                      Cap problema. Necessito les comandes de processos per veure qu\u00e8 escolta al port 10200. Executa aquestes:<\/p>\n\n\n\n

                                      1. Quin proc\u00e9s escolta al port 10200?<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1ss -tlnp | grep 10200<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      2. Processos Python\/speech en marxa:<\/strong><\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1ps aux | grep -v grep | grep -E \u00abpython|speech|piper\u00bb<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      3. Si no hi ha res en marxa<\/strong>, busca el codi en llocs habituals fora del home:<\/p>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1ls -la \/opt \/srv 2>\/dev\/null | head -40<\/p>\n\n\n\n

                                      2find \/opt \/srv \/usr\/local\/bin -maxdepth 3 -type f \\( -name \u00ab*tts*\u00bb -o -name \u00ab*speech*\u00bb \\) 2>\/dev\/null<\/p>\n\n\n\n

                                      3<\/p>\n\n\n\n

                                      3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      La comanda 1<\/strong> \u00e9s la m\u00e9s directa: et dir\u00e0 el PID i la l\u00ednia de comandes exacta del servidor que at\u00e9n les peticions de prova_piper.py<\/code>. Si el servidor no est\u00e0 en marxa ara mateix, digues-m’ho i busquem el codi per un altre cam\u00ed (p. ex. history<\/code> o nohup.out<\/code>).<\/p>\n\n\n\n

                                      Enrutamiento Inteligente<\/p>\n\n\n\n

                                      Miquel@Miquel-PC:~$ ss -tlnp | grep 10200 LISTEN 0 4096 0.0.0.0:10200 0.0.0.0:* LISTEN 0 4096 [::]:10200 [::]:* Miquel@Miquel-PC:~$ ps aux | grep -v grep | grep -E \u00abpython|speech|piper\u00bb root 1777 0.0 0.0 220 76 ? S 10:57 0:00 s6-supervise svc-piper 911 1934 0.0 0.8 288484 64444 ? Ssl 10:57 0:01 python3 -m wyoming_piper –uri tcp:\/\/0.0.0.0:10200 –length-scale 1.0 –noise-scale 0.667 –noise-w 0.333 –speaker 0 –voice ca_ES-upc_ona-medium –data-dir \/config –download-dir \/config –update-voices Miquel 5861 38.4 3.9 1186907532 316636 ? Rl 11:04 27:18 \/usr\/lib\/qt6\/libexec\/QtWebEngineProcess –type=renderer –webengine-schemes=qrc:LV –no-sandbox –disable-speech-api –enable-logging –log-level=2 –disable-databases –disable-blink-features=EyeDropperAPI,WebOTP –lang=es –num-raster-threads=2 –enable-main-frame-before-activation –renderer-client-id=3 –time-ticks-at-unix-epoch=-1789808253112640 –launch-time-ticks=439341684 –shared-files=v8_context_snapshot_data:100 –field-trial-handle=0,i,8848351394108136581,9392166901195399885,262144 –enable-features=NetworkServiceInProcess2,TracingServiceInProcess –disable-features=BackgroundFetch,EyeDropper,InstalledApp,WebOTP,WebPayments,WebUSB –variations-seed-version Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                      Deep think completado (tard\u00f3 57s)<\/p>\n\n\n\n

                                      Ara ho veig tot clar! \ud83c\udfaf El teu \u00abtts-server\u00bb \u00e9s en realitat wyoming-piper<\/strong>: el servidor de TTS de Home Assistant (protocol Wyoming) executant-se dins del contenidor piper<\/code>:<\/p>\n\n\n\n

                                      plaintext<\/p>\n\n\n\n

                                      1PID 1934: python3 -m wyoming_piper –uri tcp:\/\/0.0.0.0:10200 … –voice ca_ES-upc_ona-medium<\/p>\n\n\n\n

                                      2<\/p>\n\n\n\n

                                      2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      I el teu prova_piper.py<\/code> parla el protocol Wyoming<\/strong> (missatges synthesize<\/code>, audio-start<\/code>, audio-chunk<\/code>, audio-stop<\/code>).<\/p>\n\n\n\n

                                      Aix\u00f2 canvia la perspectiva de la integraci\u00f3: el servidor wyoming-piper no passa per docker exec --output_file -<\/code><\/strong> (fa servir Piper internament), aix\u00ed que el problema de \u00abteulada\u00bb aqu\u00ed \u00e9s probablement un altre: la puntuaci\u00f3 final de frase<\/strong>. Recorda que mai vam aclarir del tot si Piper requereix el punt final \u2014 ho hem de verificar.<\/p>\n\n\n\n

                                      Prova 1: Piper requereix el punt final? (dins el contenidor)<\/h2>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1docker exec piper bash -c \u00abecho ‘teulada’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/t1.wav && wc -c \/tmp\/t1.wav\u00bb<\/p>\n\n\n\n

                                      2docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/t2.wav && wc -c \/tmp\/t2.wav\u00bb<\/p>\n\n\n\n

                                      3<\/p>\n\n\n\n

                                      3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      Prova 2: El servidor wyoming (port 10200) respon per a textos curts?<\/h2>\n\n\n\n

                                      bash<\/p>\n\n\n\n

                                      1python3 – <<‘EOF’<\/p>\n\n\n\n

                                      2import socket, json<\/p>\n\n\n\n

                                      3<\/p>\n\n\n\n

                                      4def test(texto):<\/p>\n\n\n\n

                                      5 s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10)<\/p>\n\n\n\n

                                      6 f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                      7 evt = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: texto, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}})<\/p>\n\n\n\n

                                      8 f.write((evt + \u00ab\\n\\n\u00bb).encode()); f.flush()<\/p>\n\n\n\n

                                      9 total = 0<\/p>\n\n\n\n

                                      10 while True:<\/p>\n\n\n\n

                                      11 h = f.readline()<\/p>\n\n\n\n

                                      12 if not h: break<\/p>\n\n\n\n

                                      13 m = json.loads(h)<\/p>\n\n\n\n

                                      14 if m.get(\u00abtype\u00bb) == \u00abaudio-chunk\u00bb:<\/p>\n\n\n\n

                                      15 total += m[\u00abdata\u00bb][\u00abbytes\u00bb]; f.read(m[\u00abdata\u00bb][\u00abbytes\u00bb])<\/p>\n\n\n\n

                                      16 elif m.get(\u00abtype\u00bb) == \u00abaudio-stop\u00bb: break<\/p>\n\n\n\n

                                      17 s.close()<\/p>\n\n\n\n

                                      18 print(f\u00bb{texto!r}: {total} bytes d’\u00e0udio\u00bb)<\/p>\n\n\n\n

                                      19<\/p>\n\n\n\n

                                      20test(\u00abteulada\u00bb)<\/p>\n\n\n\n

                                      21test(\u00abteulada.\u00bb)<\/p>\n\n\n\n

                                      22EOF<\/p>\n\n\n\n

                                      23<\/p>\n\n\n\n

                                      23 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                      Qu\u00e8 esperem<\/h2>\n\n\n\n
                                        \n
                                      • Si Prova 1<\/strong>: t1<\/code> = 0 i t2<\/code> > 0 \u2192 Piper exigeix punt final<\/strong>.<\/li>\n\n\n\n
                                      • Si Prova 2<\/strong>: \u00abteulada\u00bb = 0 i \u00abteulada.\u00bb > 0 \u2192 el servidor falla per falta de puntuaci\u00f3<\/strong>, i la integraci\u00f3 consistir\u00e0 a fer que wyoming-piper afegeixi el punt autom\u00e0ticament (per exemple amb un wrapper al binari piper<\/code> dins del contenidor).<\/li>\n<\/ul>\n\n\n\n

                                        Executa les dues proves i enganxa’m els resultats.<\/p>\n\n\n\n

                                        Enrutamiento Inteligente<\/p>\n\n\n\n

                                        Miquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho ‘teulada’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/t1.wav && wc -c \/tmp\/t1.wav\u00bb docker exec piper bash -c \u00abecho ‘teulada.’ | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file \/tmp\/t2.wav && wc -c \/tmp\/t2.wav\u00bb 42540 \/tmp\/t1.wav 28204 \/tmp\/t2.wav Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                        Deep think completado (tard\u00f3 3s)<\/p>\n\n\n\n

                                        Resultat molt aclaridor! Ambdues opcions generen \u00e0udio:<\/p>\n\n\n\n

                                        Text<\/th>Fitxer<\/th>Bytes<\/th><\/tr><\/thead>
                                        teulada<\/code> (sense punt)<\/td>t1.wav<\/td>42.540<\/strong> \u2705<\/td><\/tr>
                                        teulada.<\/code> (amb punt)<\/td>t2.wav<\/td>28.204<\/strong> \u2705<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                                        Piper no requereix punt final<\/strong> \u2014 la hip\u00f2tesi de la puntuaci\u00f3 queda descartada. Els 0 bytes del principi eren nom\u00e9s el problema del stdout de docker exec<\/code>.<\/p>\n\n\n\n

                                        Ara falta la Prova 2<\/strong> (la que et vaig donar): comprovar si el servidor wyoming-piper (port 10200) respon amb \u00e0udio per a \u00abteulada\u00bb. Executa-la:<\/p>\n\n\n\n

                                        bash<\/p>\n\n\n\n

                                        1python3 – <<‘EOF’<\/p>\n\n\n\n

                                        2import socket, json<\/p>\n\n\n\n

                                        3<\/p>\n\n\n\n

                                        4def test(texto):<\/p>\n\n\n\n

                                        5 s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10)<\/p>\n\n\n\n

                                        6 f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                        7 evt = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: texto, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}})<\/p>\n\n\n\n

                                        8 f.write((evt + \u00ab\\n\\n\u00bb).encode()); f.flush()<\/p>\n\n\n\n

                                        9 total = 0<\/p>\n\n\n\n

                                        10 while True:<\/p>\n\n\n\n

                                        11 h = f.readline()<\/p>\n\n\n\n

                                        12 if not h: break<\/p>\n\n\n\n

                                        13 m = json.loads(h)<\/p>\n\n\n\n

                                        14 if m.get(\u00abtype\u00bb) == \u00abaudio-chunk\u00bb:<\/p>\n\n\n\n

                                        15 total += m[\u00abdata\u00bb][\u00abbytes\u00bb]; f.read(m[\u00abdata\u00bb][\u00abbytes\u00bb])<\/p>\n\n\n\n

                                        16 elif m.get(\u00abtype\u00bb) == \u00abaudio-stop\u00bb: break<\/p>\n\n\n\n

                                        17 s.close()<\/p>\n\n\n\n

                                        18 print(f\u00bb{texto!r}: {total} bytes d’\u00e0udio\u00bb)<\/p>\n\n\n\n

                                        19<\/p>\n\n\n\n

                                        20test(\u00abteulada\u00bb)<\/p>\n\n\n\n

                                        21test(\u00abteulada.\u00bb)<\/p>\n\n\n\n

                                        22EOF<\/p>\n\n\n\n

                                        23<\/p>\n\n\n\n

                                        23 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                        Aix\u00f2 ens dir\u00e0 si el servidor ja funciona per a textos curts (i llavors la integraci\u00f3 \u00e9s al teu client\/script) o si cal arreglar wyoming-piper.<\/p>\n\n\n\n

                                        Enrutamiento Inteligente<\/p>\n\n\n\n

                                        Miquel@Miquel-PC:~$ python3 – <<‘EOF’ import socket, json def test(texto): s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10) f = s.makefile(\u00abrwb\u00bb) evt = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: texto, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}}) f.write((evt + \u00ab\\n\\n\u00bb).encode()); f.flush() total = 0 while True: h = f.readline() if not h: break m = json.loads(h) if m.get(\u00abtype\u00bb) == \u00abaudio-chunk\u00bb: total += m[\u00abdata\u00bb][\u00abbytes\u00bb]; f.read(m[\u00abdata\u00bb][\u00abbytes\u00bb]) elif m.get(\u00abtype\u00bb) == \u00abaudio-stop\u00bb: break s.close() print(f\u00bb{texto!r}: {total} bytes d’\u00e0udio\u00bb) test(\u00abteulada\u00bb) test(\u00abteulada.\u00bb) EOF Traceback (most recent call last): File \u00ab<stdin>\u00bb, line 19, in <module> File \u00ab<stdin>\u00bb, line 12, in test File \u00ab\/usr\/lib\/python3.12\/json\/__init__.py\u00bb, line 346, in loads return _default_decoder.decode(s) ^^^^^^^^^^^^^^^^^^^^^^^^^^ File \u00ab\/usr\/lib\/python3.12\/json\/decoder.py\u00bb, line 341, in decode raise JSONDecodeError(\u00abExtra data\u00bb, s, end) json.decoder.JSONDecodeError: Extra data: line 1 column 62 (char 61) Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                        Deep think completado (tard\u00f3 33s)<\/p>\n\n\n\n

                                        El test ha fallat amb un error de JSON: \u00abExtra data: line 1 column 62\u00bb<\/strong>. Aix\u00f2 vol dir que la primera l\u00ednia que torna el servidor no \u00e9s un sol objecte JSON<\/strong> \u2014 hi ha m\u00e9s dades enganxades, o el format real de la resposta \u00e9s diferent del que assum\u00edem. Necessito veure la resposta crua<\/strong> del servidor per entendre el protocol real.<\/p>\n\n\n\n

                                        Executa aquesta depuraci\u00f3<\/h2>\n\n\n\n

                                        bash<\/p>\n\n\n\n

                                        1python3 – <<‘EOF’<\/p>\n\n\n\n

                                        2import socket, json<\/p>\n\n\n\n

                                        3<\/p>\n\n\n\n

                                        4s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10)<\/p>\n\n\n\n

                                        5f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                        6evt = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: \u00abteulada\u00bb, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}})<\/p>\n\n\n\n

                                        7print(\u00abENVIO:\u00bb, evt)<\/p>\n\n\n\n

                                        8f.write((evt + \u00ab\\n\\n\u00bb).encode()); f.flush()<\/p>\n\n\n\n

                                        9<\/p>\n\n\n\n

                                        10s.settimeout(3)<\/p>\n\n\n\n

                                        11try:<\/p>\n\n\n\n

                                        12 raw = s.recv(200)<\/p>\n\n\n\n

                                        13 print(\u00abREBUT cru:\u00bb, repr(raw))<\/p>\n\n\n\n

                                        14except socket.timeout:<\/p>\n\n\n\n

                                        15 print(\u00ab(timeout: el servidor no respon)\u00bb)<\/p>\n\n\n\n

                                        16s.close()<\/p>\n\n\n\n

                                        17EOF<\/p>\n\n\n\n

                                        18<\/p>\n\n\n\n

                                        18 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                        El repr()<\/code> et mostrar\u00e0 els bytes exactes que envia el servidor (incloent salts de l\u00ednia i car\u00e0cters de control). Amb aix\u00f2 veur\u00e9 si:<\/p>\n\n\n\n

                                          \n
                                        • El servidor espera un handshake previ<\/strong> (p. ex. un event describe<\/code> primer, com fa el protocol Wyoming est\u00e0ndard)<\/li>\n\n\n\n
                                        • O la resposta t\u00e9 un format diferent del que fa servir el teu prova_piper.py<\/code><\/li>\n<\/ul>\n\n\n\n

                                          Recorda que prova_piper.py<\/code> no l’hem vist funcionar mai en aquesta conversa<\/strong> \u2014 potser el protocol del servidor real \u00e9s diferent del que assumeix el teu client.<\/p>\n\n\n\n

                                          Enrutamiento Inteligente<\/p>\n\n\n\n

                                          Miquel@Miquel-PC:~$ python3 – <<‘EOF’ import socket, json s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10) f = s.makefile(\u00abrwb\u00bb) evt = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: \u00abteulada\u00bb, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}}) print(\u00abENVIO:\u00bb, evt) f.write((evt + \u00ab\\n\\n\u00bb).encode()); f.flush() s.settimeout(3) try: raw = s.recv(200) print(\u00abREBUT cru:\u00bb, repr(raw)) except socket.timeout: print(\u00ab(timeout: el servidor no respon)\u00bb) s.close() EOF ENVIO: {\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata\u00bb: {\u00abtext\u00bb: \u00abteulada\u00bb, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}} REBUT cru: b'{\u00abtype\u00bb: \u00abaudio-start\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61}\\n{\u00abrate\u00bb: 22050, \u00abwidth\u00bb: 2, \u00abchannels\u00bb: 1, \u00abtimestamp\u00bb: null}{\u00abtype\u00bb: \u00abaudio-chunk\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61, \u00abpayload_le’ Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                          Deep think completado (tard\u00f3 83s)<\/p>\n\n\n\n

                                          Ara s\u00ed que veig el protocol real! La resposta crua del servidor:<\/p>\n\n\n\n

                                          plaintext<\/p>\n\n\n\n

                                          1{\u00abtype\u00bb: \u00abaudio-start\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61}\\n{\u00abrate\u00bb: 22050, …}{\u00abtype\u00bb: \u00abaudio-chunk\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61, …<\/p>\n\n\n\n

                                          2<\/p>\n\n\n\n

                                          2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                          Dues conclusions importants<\/h2>\n\n\n\n

                                          1. Bona not\u00edcia: el servidor wyoming-piper S\u00cd que genera \u00e0udio per a \u00abteulada\u00bb!<\/strong> \ud83c\udf89 Ha respost amb audio-start<\/code> i audio-chunk<\/code> \u2014 el problema dels textos curts era nom\u00e9s del cam\u00ed docker exec --output_file -<\/code>, no del servidor.<\/p>\n\n\n\n

                                          2. El teu prova_piper.py<\/code> parla un protocol incorrecte.<\/strong> El protocol Wyoming real separa cap\u00e7alera<\/strong> (amb data_length<\/code>) i payload<\/strong> (els bytes seg\u00fcents). El teu client assumia que la l\u00ednia JSON contenia les dades directament (data: {bytes: N}<\/code>) \u2014 per aix\u00f2 petaria en llegir la resposta.<\/p>\n\n\n\n

                                          La integraci\u00f3: un client Wyoming correcte<\/h2>\n\n\n\n

                                          Guarda aquest script (reempla\u00e7a el protocol erroni):<\/p>\n\n\n\n

                                          bash<\/p>\n\n\n\n

                                          1cat > \/home\/Miquel\/tts-server\/tts_wyoming.py <<‘EOF’<\/p>\n\n\n\n

                                          2#!\/usr\/bin\/env python3<\/p>\n\n\n\n

                                          3″\u00bb\u00bbClient Wyoming correcte per a wyoming-piper (TTS) – port 10200.\u00bb\u00bb\u00bb<\/p>\n\n\n\n

                                          4import json, socket, subprocess, sys, wave<\/p>\n\n\n\n

                                          5<\/p>\n\n\n\n

                                          6HOST, PORT = \u00ab127.0.0.1\u00bb, 10200<\/p>\n\n\n\n

                                          7VOICE = \u00abca_ES-upc_ona-medium\u00bb<\/p>\n\n\n\n

                                          8OUT = \u00ab\/tmp\/veu_wyoming.wav\u00bb<\/p>\n\n\n\n

                                          9<\/p>\n\n\n\n

                                          10def send_event(f, evt_type, data=None):<\/p>\n\n\n\n

                                          11 payload = json.dumps(data or {}).encode()<\/p>\n\n\n\n

                                          12 header = json.dumps({\u00abtype\u00bb: evt_type, \u00abdata_length\u00bb: len(payload)}).encode() + b\u00bb\\n\u00bb<\/p>\n\n\n\n

                                          13 f.write(header + payload)<\/p>\n\n\n\n

                                          14 f.flush()<\/p>\n\n\n\n

                                          15<\/p>\n\n\n\n

                                          16def read_event(f):<\/p>\n\n\n\n

                                          17 line = f.readline()<\/p>\n\n\n\n

                                          18 if not line:<\/p>\n\n\n\n

                                          19 return None, None<\/p>\n\n\n\n

                                          20 meta = json.loads(line)<\/p>\n\n\n\n

                                          21 n = meta.get(\u00abdata_length\u00bb, 0)<\/p>\n\n\n\n

                                          22 payload = f.read(n) if n > 0 else b\u00bb\u00bb<\/p>\n\n\n\n

                                          23 return meta[\u00abtype\u00bb], payload<\/p>\n\n\n\n

                                          24<\/p>\n\n\n\n

                                          25def main():<\/p>\n\n\n\n

                                          26 text = sys.argv[1] if len(sys.argv) > 1 else \u00abHola\u00bb<\/p>\n\n\n\n

                                          27 s = socket.create_connection((HOST, PORT), timeout=30)<\/p>\n\n\n\n

                                          28 f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                          29 send_event(f, \u00absynthesize\u00bb, {\u00abtext\u00bb: text, \u00abvoice\u00bb: VOICE})<\/p>\n\n\n\n

                                          30 chunks, rate, width, channels = [], 22050, 2, 1<\/p>\n\n\n\n

                                          31 while True:<\/p>\n\n\n\n

                                          32 t, payload = read_event(f)<\/p>\n\n\n\n

                                          33 if t is None:<\/p>\n\n\n\n

                                          34 print(\u00abConnexi\u00f3 tancada pel servidor\u00bb); break<\/p>\n\n\n\n

                                          35 if t == \u00abaudio-start\u00bb:<\/p>\n\n\n\n

                                          36 info = json.loads(payload)<\/p>\n\n\n\n

                                          37 rate = info.get(\u00abrate\u00bb, rate)<\/p>\n\n\n\n

                                          38 width = info.get(\u00abwidth\u00bb, width)<\/p>\n\n\n\n

                                          39 channels = info.get(\u00abchannels\u00bb, channels)<\/p>\n\n\n\n

                                          40 elif t == \u00abaudio-chunk\u00bb:<\/p>\n\n\n\n

                                          41 chunks.append(payload)<\/p>\n\n\n\n

                                          42 elif t == \u00abaudio-stop\u00bb:<\/p>\n\n\n\n

                                          43 break<\/p>\n\n\n\n

                                          44 s.close()<\/p>\n\n\n\n

                                          45 audio = b\u00bb\u00bb.join(chunks)<\/p>\n\n\n\n

                                          46 if not audio:<\/p>\n\n\n\n

                                          47 print(\u00abERROR: no s’ha rebut \u00e0udio\u00bb); sys.exit(1)<\/p>\n\n\n\n

                                          48 with wave.open(OUT, \u00abwb\u00bb) as w:<\/p>\n\n\n\n

                                          49 w.setnchannels(channels); w.setsampwidth(width); w.setframerate(rate)<\/p>\n\n\n\n

                                          50 w.writeframes(audio)<\/p>\n\n\n\n

                                          51 print(f\u00bbOK: {OUT} ({len(audio)} bytes, {rate} Hz)\u00bb)<\/p>\n\n\n\n

                                          52 subprocess.run([\u00abaplay\u00bb, OUT])<\/p>\n\n\n\n

                                          53<\/p>\n\n\n\n

                                          54if __name__ == \u00ab__main__\u00bb:<\/p>\n\n\n\n

                                          55 main()<\/p>\n\n\n\n

                                          56EOF<\/p>\n\n\n\n

                                          57<\/p>\n\n\n\n

                                          57 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                          Prova-ho<\/h2>\n\n\n\n

                                          bash<\/p>\n\n\n\n

                                          1python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abteulada\u00bb<\/p>\n\n\n\n

                                          2python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\u00bb<\/p>\n\n\n\n

                                          3<\/p>\n\n\n\n

                                          3 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                          Aquest client parla el protocol real de wyoming-piper: envia synthesize<\/code> amb cap\u00e7alera correcta (data_length<\/code>), llegeix audio-start<\/code> \u2192 audio-chunk<\/code> \u2192 audio-stop<\/code>, guarda el WAV i el reprodueix amb aplay<\/code>. El servidor ja fa la seva feina; el que faltava era el client correcte.<\/strong> Prova-ho i digue’m si l’Ona parla tant els textos curts com els llargs!<\/p>\n\n\n\n

                                          Enrutamiento Inteligente<\/p>\n\n\n\n

                                          \"\"\/<\/figure>\n\n\n\n

                                          aix\u00ed es pot esborrar prova_piper.py?<\/p>\n\n\n\n

                                          Deep think completado (tard\u00f3 6s)<\/p>\n\n\n\n

                                          S\u00ed! Pots esborrar-los tranquil\u00b7lament \u2014 tts_wyoming.py<\/code> els substitueix:<\/p>\n\n\n\n

                                            \n
                                          • prova_piper.py<\/code><\/strong> \u2014 protocol Wyoming incorrecte (assumia data: {bytes: N}<\/code> dins la cap\u00e7alera) \u2192 obsolet.<\/li>\n\n\n\n
                                          • prova_piper_dbg.py<\/code><\/strong> \u2014 mateix protocol incorrecte, nom\u00e9s amb logs \u2192 tamb\u00e9 obsolet.<\/li>\n<\/ul>\n\n\n\n

                                            Per seguretat, en comptes d’esborrar-los, els moc a una carpeta de c\u00f2pia de seguretat (aix\u00ed si algun dia els necessites, hi s\u00f3n):<\/p>\n\n\n\n

                                            Deep think completado (tard\u00f3 1s)<\/p>\n\n\n\n

                                            Fet! \u2705 Els dos scripts antics han quedat moguts a:<\/p>\n\n\n\n

                                            plaintext<\/p>\n\n\n\n

                                            1\/home\/Miquel\/tts-server\/backup_scripts\/<\/p>\n\n\n\n

                                            2\u251c\u2500\u2500 prova_piper.py<\/p>\n\n\n\n

                                            3\u2514\u2500\u2500 prova_piper_dbg.py<\/p>\n\n\n\n

                                            4<\/p>\n\n\n\n

                                            4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                            Estructura final de tts-server\/<\/code><\/h2>\n\n\n\n
                                            Fitxer\/carpeta<\/th>Estat<\/th><\/tr><\/thead>
                                            tts_wyoming.py<\/code><\/strong><\/td>\u2705 El client correcte (protocol Wyoming real)<\/td><\/tr>
                                            piper-config\/<\/code><\/td>Models de Piper<\/td><\/tr>
                                            tts-server\/<\/code><\/td>Subcarpeta amb c\u00f2pies<\/td><\/tr>
                                            venv\/<\/code><\/td>Entorn Python<\/td><\/tr>
                                            server.log<\/code><\/td>Logs<\/td><\/tr>
                                            backup_scripts\/<\/code><\/td>Antics scripts (obsolets)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

                                            Si ho vols netejar del tot, ara pots esborrar la carpeta backup_scripts<\/code> amb rm -r<\/code> \u2014 o deixar-la com a c\u00f2pia de seguretat.<\/p>\n\n\n\n

                                            Recorda que l’\u00fanic que necessites fer servir d’ara endavant \u00e9s:<\/p>\n\n\n\n

                                            bash<\/p>\n\n\n\n

                                            1python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abteulada\u00bb<\/p>\n\n\n\n

                                            2<\/p>\n\n\n\n

                                            2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                            I si algun dia vols fer-ho servir com a ordre r\u00e0pida, pots crear un enlla\u00e7 o un alias<\/code>:<\/p>\n\n\n\n

                                            bash<\/p>\n\n\n\n

                                            1echo ‘alias ona=\u00bbpython3 \/home\/Miquel\/tts-server\/tts_wyoming.py\u00bb‘ >> ~\/.bashrc<\/p>\n\n\n\n

                                            2source ~\/.bashrc<\/p>\n\n\n\n

                                            3ona \u00abHola, s\u00f3c l’Ona!\u00bb<\/p>\n\n\n\n

                                            4<\/p>\n\n\n\n

                                            4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                            Vols que t’ajudi amb alguna altra cosa?<\/p>\n\n\n\n

                                            Enrutamiento Inteligente<\/p>\n\n\n\n

                                            fem una prova amb un text cur, com ‘teulada’<\/p>\n\n\n\n

                                            Deep think completado (tard\u00f3 4s)<\/p>\n\n\n\n

                                            Endavant! Executa aix\u00f2 al teu terminal:<\/p>\n\n\n\n

                                            bash<\/p>\n\n\n\n

                                            1python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abteulada\u00bb<\/p>\n\n\n\n

                                            2<\/p>\n\n\n\n

                                            2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                            Qu\u00e8 has d’esperar<\/h2>\n\n\n\n
                                              \n
                                            1. El script connecta a wyoming-piper (port 10200) i envia synthesize<\/code> amb el text teulada<\/code>.<\/li>\n\n\n\n
                                            2. El servidor sintetitza (recorda que ja vam veure que respon amb audio-start<\/code> + audio-chunk<\/code> per a textos curts).<\/li>\n\n\n\n
                                            3. Veur\u00e0s un missatge com: plaintext 1OK: \/tmp\/veu_wyoming.wav (XXXXX bytes, 22050 Hz)22 l\u00edneas de c\u00f3digo colapsadasExpandir<\/li>\n\n\n\n
                                            4. Llavors aplay<\/code> reproduir\u00e0 el WAV i hauries de sentir \u00abteulada\u00bb<\/strong> \ud83d\udd0a<\/li>\n<\/ol>\n\n\n\n

                                              Si surt el OK<\/code> i sona, ja tenim la integraci\u00f3 completa funcionant amb textos curts. Digue’m qu\u00e8 passa!<\/p>\n\n\n\n

                                              Enrutamiento Inteligente<\/p>\n\n\n\n

                                              Miquel@Miquel-PC:~$ python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abteulada\u00bb Traceback (most recent call last): File \u00ab\/home\/Miquel\/tts-server\/tts_wyoming.py\u00bb, line 54, in <module> main() File \u00ab\/home\/Miquel\/tts-server\/tts_wyoming.py\u00bb, line 31, in main t, payload = read_event(f) ^^^^^^^^^^^^^ File \u00ab\/home\/Miquel\/tts-server\/tts_wyoming.py\u00bb, line 19, in read_event meta = json.loads(line) ^^^^^^^^^^^^^^^^ File \u00ab\/usr\/lib\/python3.12\/json\/__init__.py\u00bb, line 341, in loads s = s.decode(detect_encoding(s), ‘surrogatepass’) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File \u00ab\/usr\/lib\/python3.12\/encodings\/utf_16_le.py\u00bb, line 16, in decode return codecs.utf_16_le_decode(input, errors, True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ UnicodeDecodeError: ‘utf-16-le’ codec can’t decode byte 0x0a in position 444: truncated data decoding with ‘utf-16-le’ codec failed Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                              Deep think completado (tard\u00f3 47s)<\/p>\n\n\n\n

                                              L’error ens dona una pista clau: json.loads<\/code> ha rebut una l\u00ednia de m\u00e9s de 444 bytes plena de binaris (per aix\u00f2 detecta UTF-16). Aix\u00f2 vol dir que la lectura s’ha desincronitzat<\/strong>: hem llegit \u00e0udio cru com si fos una cap\u00e7alera. La meva suposici\u00f3 sobre el camp data_length<\/code> dels audio-chunk<\/code> deu ser incorrecta.<\/p>\n\n\n\n

                                              Necessito veure la seq\u00fc\u00e8ncia completa d’esdeveniments. Executa aquesta depuraci\u00f3 que mostra cada cap\u00e7alera i el seu payload:<\/p>\n\n\n\n

                                              bash<\/p>\n\n\n\n

                                              1python3 – <<‘EOF’<\/p>\n\n\n\n

                                              2import socket, json<\/p>\n\n\n\n

                                              3<\/p>\n\n\n\n

                                              4s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10)<\/p>\n\n\n\n

                                              5f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                              6<\/p>\n\n\n\n

                                              7# Envia synthesize amb el protocol correcte<\/p>\n\n\n\n

                                              8payload = json.dumps({\u00abtext\u00bb: \u00abteulada\u00bb, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}).encode()<\/p>\n\n\n\n

                                              9header = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata_length\u00bb: len(payload)}).encode() + b\u00bb\\n\u00bb<\/p>\n\n\n\n

                                              10f.write(header + payload); f.flush()<\/p>\n\n\n\n

                                              11<\/p>\n\n\n\n

                                              12for i in range(8):<\/p>\n\n\n\n

                                              13 line = f.readline()<\/p>\n\n\n\n

                                              14 if not line:<\/p>\n\n\n\n

                                              15 print(f\u00bb— event {i}: fi de connexi\u00f3\u00bb); break<\/p>\n\n\n\n

                                              16 print(f\u00bb— event {i} —\u00ab)<\/p>\n\n\n\n

                                              17 print(\u00abheader:\u00bb, repr(line[:120]))<\/p>\n\n\n\n

                                              18 try:<\/p>\n\n\n\n

                                              19 meta = json.loads(line)<\/p>\n\n\n\n

                                              20 n = meta.get(\u00abdata_length\u00bb, 0)<\/p>\n\n\n\n

                                              21 print(\u00abtype:\u00bb, meta.get(\u00abtype\u00bb), \u00ab| data_length:\u00bb, n, \u00ab| claus:\u00bb, list(meta.keys()))<\/p>\n\n\n\n

                                              22 body = f.read(n)<\/p>\n\n\n\n

                                              23 print(\u00abpayload:\u00bb, repr(body[:60]), \u00ab… len:\u00bb, len(body))<\/p>\n\n\n\n

                                              24 except Exception as e:<\/p>\n\n\n\n

                                              25 print(\u00abERROR:\u00bb, e)<\/p>\n\n\n\n

                                              26 break<\/p>\n\n\n\n

                                              27s.close()<\/p>\n\n\n\n

                                              28EOF<\/p>\n\n\n\n

                                              29<\/p>\n\n\n\n

                                              29 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                              Aix\u00f2 em mostrar\u00e0:<\/p>\n\n\n\n

                                                \n
                                              • Quants bytes declara cada audio-chunk<\/code> (data_length<\/code>)<\/li>\n\n\n\n
                                              • Si hi ha un camp addicional (com payload_length<\/code>) que \u00e9s el que realment indica la mida de l’\u00e0udio<\/li>\n\n\n\n
                                              • On exactament es desincronitza la lectura<\/li>\n<\/ul>\n\n\n\n

                                                Amb aix\u00f2 corregeixo el client perqu\u00e8 llegeixi el protocol exacte de wyoming-piper.<\/p>\n\n\n\n

                                                Enrutamiento Inteligente<\/p>\n\n\n\n

                                                Miquel@Miquel-PC:~$ python3 – <<‘EOF’ import socket, json s = socket.create_connection((\u00ab127.0.0.1\u00bb, 10200), timeout=10) f = s.makefile(\u00abrwb\u00bb) # Envia synthesize amb el protocol correcte payload = json.dumps({\u00abtext\u00bb: \u00abteulada\u00bb, \u00abvoice\u00bb: \u00abca_ES-upc_ona-medium\u00bb}).encode() header = json.dumps({\u00abtype\u00bb: \u00absynthesize\u00bb, \u00abdata_length\u00bb: len(payload)}).encode() + b\u00bb\\n\u00bb f.write(header + payload); f.flush() for i in range(8): line = f.readline() if not line: print(f\u00bb— event {i}: fi de connexi\u00f3\u00bb); break print(f\u00bb— event {i} —\u00ab) print(\u00abheader:\u00bb, repr(line[:120])) try: meta = json.loads(line) n = meta.get(\u00abdata_length\u00bb, 0) print(\u00abtype:\u00bb, meta.get(\u00abtype\u00bb), \u00ab| data_length:\u00bb, n, \u00ab| claus:\u00bb, list(meta.keys())) body = f.read(n) print(\u00abpayload:\u00bb, repr(body[:60]), \u00ab… len:\u00bb, len(body)) except Exception as e: print(\u00abERROR:\u00bb, e) break s.close() EOF — event 0 — header: b'{\u00abtype\u00bb: \u00abaudio-start\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61}\\n’ type: audio-start | data_length: 61 | claus: [‘type’, ‘version’, ‘data_length’] payload: b'{\u00abrate\u00bb: 22050, \u00abwidth\u00bb: 2, \u00abchannels\u00bb: 1, \u00abtimestamp\u00bb: null’ … len: 61 — event 1 — header: b'{\u00abtype\u00bb: \u00abaudio-chunk\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61, \u00abpayload_length\u00bb: 2048}\\n’ type: audio-chunk | data_length: 61 | claus: [‘type’, ‘version’, ‘data_length’, ‘payload_length’] payload: b'{\u00abrate\u00bb: 22050, \u00abwidth\u00bb: 2, \u00abchannels\u00bb: 1, \u00abtimestamp\u00bb: null’ … len: 61 — event 2 — header: b’@\\x01y\\x00\\xca\\xff\\xe5\\xfer\\xfe\\xdb\\xfed\\xfe<\\xff\\x89\\xfe\\xaf\\xfe\\xfb\\xfe\\xb6\\xff\\xa9\\xffA\\x00\\x80\\x00″\\x01v\\x01N\\x01\\x04\\x02K\\x01,\\x02\\xcc\\x01R\\x01\\xa1\\x01\\x9d\\x01\\xc4\\x01&\\x02z\\x02\\x0c\\x02\\xb9\\x02M\\x02\\x14\\x02|\\x01\\xfb\\x01\\xb5\\x02\\x0f\\x01\\xe3\\x02&\\x02\\x07\\x03\\x16\\x03;\\x02\\xeb\\x01y\\x00b\\x03\\xf8\\x00\\xfa\\x01\\x7f\\x02\\x8a\\x01\\xf0\\x02\\x0e\\x02\\xc9\\x03\\x80\\x03t\\x03\\x92\\x03q\\x01\/\\x03\\x9d\\x01\\x86\\x02\\xa9\\x02V\\x03′ ERROR: ‘utf-8’ codec can’t decode byte 0xca in position 4: invalid continuation byte Miquel@Miquel-PC:~$<\/p>\n\n\n\n

                                                Deep think completado (tard\u00f3 32s)<\/p>\n\n\n\n

                                                Ja ho veig! La depuraci\u00f3 ha revelat el format real dels audio-chunk<\/code> de wyoming-piper:<\/p>\n\n\n\n

                                                plaintext<\/p>\n\n\n\n

                                                1{\u00abtype\u00bb: \u00abaudio-chunk\u00bb, \u00abversion\u00bb: \u00ab1.10.2\u00bb, \u00abdata_length\u00bb: 61, \u00abpayload_length\u00bb: 2048}<\/p>\n\n\n\n

                                                2\u251c\u2500\u2500 data_length (61) \u2192 JSON de metadades (rate\/width\/channels\/timestamp)<\/p>\n\n\n\n

                                                3\u2514\u2500\u2500 payload_length (2048) \u2192 L’\u00c0UDIO REAL<\/p>\n\n\n\n

                                                4<\/p>\n\n\n\n

                                                4 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                                El meu client nom\u00e9s llegia data_length<\/code> (61 bytes de metadades) i ignorava els 2048 bytes d’\u00e0udio<\/strong> que venen despr\u00e9s \u2014 per aix\u00f2 la lectura es desincronitzava i llegia \u00e0udio cru com a cap\u00e7alera.<\/p>\n\n\n\n

                                                Corregeix el client<\/h2>\n\n\n\n

                                                bash<\/p>\n\n\n\n

                                                1cat > \/home\/Miquel\/tts-server\/tts_wyoming.py <<‘EOF’<\/p>\n\n\n\n

                                                2#!\/usr\/bin\/env python3<\/p>\n\n\n\n

                                                3″\u00bb\u00bbClient Wyoming correcte per a wyoming-piper (TTS) – port 10200.\u00bb\u00bb\u00bb<\/p>\n\n\n\n

                                                4import json, socket, subprocess, sys, wave<\/p>\n\n\n\n

                                                5<\/p>\n\n\n\n

                                                6HOST, PORT = \u00ab127.0.0.1\u00bb, 10200<\/p>\n\n\n\n

                                                7VOICE = \u00abca_ES-upc_ona-medium\u00bb<\/p>\n\n\n\n

                                                8OUT = \u00ab\/tmp\/veu_wyoming.wav\u00bb<\/p>\n\n\n\n

                                                9<\/p>\n\n\n\n

                                                10def send_event(f, evt_type, data=None):<\/p>\n\n\n\n

                                                11 payload = json.dumps(data or {}).encode()<\/p>\n\n\n\n

                                                12 header = json.dumps({\u00abtype\u00bb: evt_type, \u00abdata_length\u00bb: len(payload)}).encode() + b\u00bb\\n\u00bb<\/p>\n\n\n\n

                                                13 f.write(header + payload)<\/p>\n\n\n\n

                                                14 f.flush()<\/p>\n\n\n\n

                                                15<\/p>\n\n\n\n

                                                16def read_event(f):<\/p>\n\n\n\n

                                                17 line = f.readline()<\/p>\n\n\n\n

                                                18 if not line:<\/p>\n\n\n\n

                                                19 return None, None, None<\/p>\n\n\n\n

                                                20 meta = json.loads(line)<\/p>\n\n\n\n

                                                21 n = meta.get(\u00abdata_length\u00bb, 0)<\/p>\n\n\n\n

                                                22 data = f.read(n) if n > 0 else b\u00bb\u00bb<\/p>\n\n\n\n

                                                23 pn = meta.get(\u00abpayload_length\u00bb, 0) # camp extra amb l’\u00e0udio real<\/p>\n\n\n\n

                                                24 payload = f.read(pn) if pn > 0 else b\u00bb\u00bb<\/p>\n\n\n\n

                                                25 return meta[\u00abtype\u00bb], data, payload<\/p>\n\n\n\n

                                                26<\/p>\n\n\n\n

                                                27def main():<\/p>\n\n\n\n

                                                28 text = sys.argv[1] if len(sys.argv) > 1 else \u00abHola\u00bb<\/p>\n\n\n\n

                                                29 s = socket.create_connection((HOST, PORT), timeout=30)<\/p>\n\n\n\n

                                                30 f = s.makefile(\u00abrwb\u00bb)<\/p>\n\n\n\n

                                                31 send_event(f, \u00absynthesize\u00bb, {\u00abtext\u00bb: text, \u00abvoice\u00bb: VOICE})<\/p>\n\n\n\n

                                                32 chunks, rate, width, channels = [], 22050, 2, 1<\/p>\n\n\n\n

                                                33 while True:<\/p>\n\n\n\n

                                                34 t, data, payload = read_event(f)<\/p>\n\n\n\n

                                                35 if t is None:<\/p>\n\n\n\n

                                                36 print(\u00abConnexi\u00f3 tancada pel servidor\u00bb); break<\/p>\n\n\n\n

                                                37 if t == \u00abaudio-start\u00bb:<\/p>\n\n\n\n

                                                38 info = json.loads(data)<\/p>\n\n\n\n

                                                39 rate = info.get(\u00abrate\u00bb, rate)<\/p>\n\n\n\n

                                                40 width = info.get(\u00abwidth\u00bb, width)<\/p>\n\n\n\n

                                                41 channels = info.get(\u00abchannels\u00bb, channels)<\/p>\n\n\n\n

                                                42 elif t == \u00abaudio-chunk\u00bb:<\/p>\n\n\n\n

                                                43 chunks.append(payload) # nom\u00e9s l’\u00e0udio real<\/p>\n\n\n\n

                                                44 elif t == \u00abaudio-stop\u00bb:<\/p>\n\n\n\n

                                                45 break<\/p>\n\n\n\n

                                                46 s.close()<\/p>\n\n\n\n

                                                47 audio = b\u00bb\u00bb.join(chunks)<\/p>\n\n\n\n

                                                48 if not audio:<\/p>\n\n\n\n

                                                49 print(\u00abERROR: no s’ha rebut \u00e0udio\u00bb); sys.exit(1)<\/p>\n\n\n\n

                                                50 with wave.open(OUT, \u00abwb\u00bb) as w:<\/p>\n\n\n\n

                                                51 w.setnchannels(channels); w.setsampwidth(width); w.setframerate(rate)<\/p>\n\n\n\n

                                                52 w.writeframes(audio)<\/p>\n\n\n\n

                                                53 print(f\u00bbOK: {OUT} ({len(audio)} bytes, {rate} Hz)\u00bb)<\/p>\n\n\n\n

                                                54 subprocess.run([\u00abaplay\u00bb, OUT])<\/p>\n\n\n\n

                                                55<\/p>\n\n\n\n

                                                56if __name__ == \u00ab__main__\u00bb:<\/p>\n\n\n\n

                                                57 main()<\/p>\n\n\n\n

                                                58EOF<\/p>\n\n\n\n

                                                59<\/p>\n\n\n\n

                                                59 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                                Prova-ho<\/h2>\n\n\n\n

                                                bash<\/p>\n\n\n\n

                                                1python3 \/home\/Miquel\/tts-server\/tts_wyoming.py \u00abteulada\u00bb<\/p>\n\n\n\n

                                                2<\/p>\n\n\n\n

                                                2 l\u00edneas de c\u00f3digo colapsadasExpandir<\/p>\n\n\n\n

                                                Ara hauria de llegir les metadades (61 bytes), acumular l’\u00e0udio real (payload_length<\/code>), i reproduir \u00abteulada\u00bb correctament. Digue’m com va! \ud83d\udd0a<\/p>\n","protected":false},"excerpt":{"rendered":"

                                                UOS AI iquel@Miquel-PC:~$ docker exec piper bash -c \u00abecho \\\u00bbHola, s\u00f3c l’Ona, la meva veu s’ha desenvolupat a la UPC de Barcelona.\\\u00bb | piper –model \/config\/ca_ES-upc_ona-medium.onnx –output_file -\u00bb | aplay -f S16_LE -r 22050 -c 1 Playing WAVE ‘stdin’ : … Continua llegint →<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":3167,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[207,1],"tags":[],"class_list":["post-19155","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-deepin-linux","category-bloc-de-notes"],"_links":{"self":[{"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/posts\/19155","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.beseit.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=19155"}],"version-history":[{"count":1,"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/posts\/19155\/revisions"}],"predecessor-version":[{"id":19156,"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/posts\/19155\/revisions\/19156"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.beseit.net\/index.php?rest_route=\/wp\/v2\/media\/3167"}],"wp:attachment":[{"href":"https:\/\/www.beseit.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=19155"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.beseit.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=19155"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.beseit.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=19155"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}