feat: add fishaudio (Fish Audio) TTS support (#159)

Adds synthFishaudio with both arms: the say: streaming url consumed by the
mediajam dialect, and a POST /v1/tts cache render. The render asks for raw pcm
at 8k and returns extension r8 because fish's wav output carries a placeholder
RIFF size, the same problem gradium has.

Fish is a voice-cloning vendor, so the voice is a reference_id; the sentinel
'default' means send none and use fish's own default voice.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

lib/synth-audio.js

@@ -81,7 +81,7 @@ async function synthAudio(client, createHash, retrieveHash, logger, stats, { acc
 
   assert.ok(['google', 'aws', 'polly', 'microsoft', 'wellsaid', 'nvidia', 'elevenlabs',
     'whisper', 'deepgram', 'deepgramflux', 'rimelabs', 'cartesia', 'gradium', 'nineninesix', 'inworld', 'resemble',
-    'murf', 'xai']
+    'murf', 'xai', 'fishaudio']
     .includes(vendor) ||
   vendor.startsWith('custom'),
   `synthAudio supported vendors are google, aws, microsoft, nvidia and wellsaid ..etc, not ${vendor}`);
@@ -149,6 +149,10 @@ async function synthAudio(client, createHash, retrieveHash, logger, stats, { acc
   } else if (vendor === 'resemble') {
     assert.ok(voice, 'synthAudio requires voice when resemble is used');
     assert.ok(credentials.api_key, 'synthAudio requires api_key when resemble is used');
+  } else if ('fishaudio' === vendor) {
+    /* no voice assert: fish synthesizes with its own default voice when
+       reference_id is omitted, which is what the 'default' selection means */
+    assert.ok(credentials.api_key, 'synthAudio requires api_key when fishaudio is used');
   }
 
   const key = makeSynthKey({
@@ -220,6 +224,11 @@ async function synthAudio(client, createHash, retrieveHash, logger, stats, { acc
           credentials, options, stats, language, voice, key, text, renderForCaching, disableTtsStreaming,
           disableTtsCache});
         break;
+      case 'fishaudio':
+        audioData = await synthFishaudio(logger, {
+          credentials, options, stats, voice, key, text, renderForCaching, disableTtsStreaming,
+          disableTtsCache});
+        break;
       case 'gradium':
         audioData = await synthGradium(logger, {
           credentials, options, stats, voice, key, text, renderForCaching, disableTtsStreaming,
@@ -1505,6 +1514,81 @@ const synthGradium = async(logger, {
 
 /* nineninesix.ai — a Cartesia-compatible API, but only raw/wav come back
    (mp3 is rejected), so the cache render asks for wav rather than mp3. */
+/* fish.audio — msgpack websocket for streaming, and a POST endpoint for the cache
+   render. format:pcm + sample_rate:8000 returns bare little-endian 16-bit samples,
+   which is exactly the r8 container. we avoid fish's wav output because its RIFF
+   header carries a placeholder size (0xffffff24) — length is unknown up front, as
+   with gradium.
+
+   fish is a voice-cloning vendor: the "voice" is a reference_id returned by
+   POST /model, and omitting it entirely synthesizes with fish's default voice.
+   the sentinel value 'default' (the bundled fallback entry in the portal) means
+   exactly that — send no reference_id.
+*/
+const synthFishaudio = async(logger, {
+  credentials, options, stats, voice, key, text, renderForCaching, disableTtsStreaming, disableTtsCache
+}) => {
+  const {api_key, model_id, fishaudio_tts_uri} = credentials;
+  const {reference_id, latency, chunk_length, speed, volume} = options || {};
+
+  /* free-text reference_id in the vendor options wins over the voice selector */
+  const refId = reference_id || (voice && voice !== 'default' ? voice : null);
+
+  /* default to using the streaming interface, unless disabled by env var OR we want just a cache file */
+  if (!JAMBONES_DISABLE_TTS_STREAMING && !renderForCaching && !disableTtsStreaming) {
+    let params = '{';
+    params += `api_key=${api_key}`;
+    params += `,playback_id=${key}`;
+    params += ',vendor=fishaudio';
+    params += `,voice=${refId || 'default'}`;
+    params += `,write_cache_file=${disableTtsCache ? 0 : 1}`;
+    if (model_id) params += `,model_id=${model_id}`;
+    if (latency) params += `,latency=${latency}`;
+    if (chunk_length) params += `,chunk_length=${chunk_length}`;
+    if (speed) params += `,speed=${speed}`;
+    if (volume) params += `,volume=${volume}`;
+    if (fishaudio_tts_uri) params += `,endpoint=${fishaudio_tts_uri}`;
+    params += '}';
+
+    return {
+      filePath: `say:${params}${text.replace(/\n/g, ' ').replace(/\r/g, ' ')}`,
+      servedFromCache: false,
+      rtt: 0
+    };
+  }
+
+  try {
+    const sampleRate = 8000;
+    const post = bent(fishaudio_tts_uri || 'https://api.fish.audio', 'POST', 'buffer', {
+      'Authorization': `Bearer ${api_key}`,
+      'Content-Type': 'application/json',
+      /* the model is selected by header, not in the body */
+      'model': model_id || 's2.1-pro'
+    });
+    const audioContent = await post('/v1/tts', {
+      text,
+      format: 'pcm',
+      sample_rate: sampleRate,
+      ...(refId && {reference_id: refId}),
+      ...(latency && {latency}),
+      ...(chunk_length && {chunk_length: parseInt(chunk_length, 10)}),
+      ...((speed || volume) && {prosody: {
+        ...(speed && {speed: parseFloat(speed)}),
+        ...(volume && {volume: parseFloat(volume)})
+      }})
+    });
+    return {
+      audioContent,
+      extension: 'r8',
+      sampleRate
+    };
+  } catch (err) {
+    logger.info({err}, 'synth fishaudio returned error');
+    stats.increment('tts.count', ['vendor:fishaudio', 'accepted:no']);
+    throw err;
+  }
+};
+
 const synthNineninesix = async(logger, {
   credentials, stats, voice, language, key, text, renderForCaching, disableTtsStreaming, disableTtsCache
 }) => {

test/synth.js

@@ -1087,6 +1087,47 @@ test('gradium speech synth tests', async(t) => {
   client.quit();
 });
 
+test('fishaudio speech synth tests', async(t) => {
+  const fn = require('..');
+  const {synthAudio, client} = fn(opts, logger);
+
+  if (!process.env.FISHAUDIO_API_KEY) {
+    t.pass('skipping fishaudio speech synth tests since FISHAUDIO_API_KEY is not provided');
+    return t.end();
+  }
+  const text = 'Hi there and welcome to jambones! ' + Date.now();
+  try {
+    /* voice 'default' means "send no reference_id" — fish's own default voice */
+    const o = await synthAudio(stats, {
+      vendor: 'fishaudio',
+      credentials: {
+        api_key: process.env.FISHAUDIO_API_KEY,
+        model_id: 's2.1-pro'
+      },
+      voice: 'default',
+      text,
+      renderForCaching: true
+    });
+    t.ok(!o.servedFromCache, `successfully synthed fishaudio audio to ${o.filePath}`);
+
+    /* the cache render must be raw 8k pcm (r8): fish's wav header carries a
+       placeholder RIFF size, so we never ask for wav */
+    const o2 = await synthAudio(stats, {
+      vendor: 'fishaudio',
+      credentials: {api_key: process.env.FISHAUDIO_API_KEY},
+      voice: 'default',
+      text: text + ' two',
+      renderForCaching: true,
+      disableTtsCache: true
+    });
+    t.ok(!o2.servedFromCache, 'fishaudio synthed a second uncached render');
+  } catch (err) {
+    console.error(JSON.stringify(err));
+    t.end(err);
+  }
+  client.quit();
+});
+
 test('nineninesix speech synth tests', async(t) => {
   const fn = require('..');
   const {synthAudio, client} = fn(opts, logger);