feat(gateway-chat): repoint TTS to zonos-gateway (OpenAI /v1/audio/speech)
Point the in-page TTS at the zonos-gateway wrapper on irv-ml1:8890 (direct, so streaming isn't buffered by LiteLLM): OpenAI-shape body (model: ext-tts, input, voice), Cora default voice, float32@44.1kHz PCM decode. Quotes are joined into a single stream call (prosody — no per-sentence chunking). Migration regex rewrites stale saved endpoints (:8299/:8210, /tts[/stream], :4000) to the new one.
This commit is contained in:
+29
-19
@@ -67,11 +67,11 @@
|
||||
<div><label>Temperature</label><input id="temp" type="number" step="0.05" value="1"></div>
|
||||
<div><label>Max tokens</label><input id="max" type="number" step="1" value="1024"></div>
|
||||
</div>
|
||||
<label style="margin-top:.5rem">🔊 mOrpheus voice (quoted text)</label>
|
||||
<label style="margin-top:.5rem">🔊 ext-tts voice (quoted text)</label>
|
||||
<label class="hint" style="display:flex;gap:.45rem;align-items:center;margin:.15rem 0"><input id="ttsOn" type="checkbox" style="width:auto"> Auto-play quoted dialogue</label>
|
||||
<input id="ttsUrl" placeholder="http://10.100.79.3:8299/tts">
|
||||
<input id="ttsVoice" placeholder="baddy">
|
||||
<div class="hint">Settings persist locally. No <code>tools</code> field is ever sent. Reasoning models show their <code>reasoning_content</code> above the answer. With 🔊 on, quoted text in each reply auto-plays via mOrpheus (irv-ml1, internal).</div>
|
||||
<input id="ttsUrl" placeholder="http://10.100.79.3:8890/v1/audio/speech">
|
||||
<input id="ttsVoice" placeholder="Cora">
|
||||
<div class="hint">Settings persist locally. No <code>tools</code> field is ever sent. Reasoning models show their <code>reasoning_content</code> above the answer. With 🔊 on, quoted text in each reply auto-plays via <code>ext-tts</code> (Zonos, streamed direct from zonos-gateway).</div>
|
||||
</aside>
|
||||
<main id="main">
|
||||
<div id="log"></div>
|
||||
@@ -88,8 +88,14 @@ const $ = id => document.getElementById(id);
|
||||
const cfg = ['base','key','model','sys','temp','max','ttsUrl','ttsVoice'];
|
||||
cfg.forEach(k => { const v = localStorage.getItem('gc_'+k); if (v !== null) $(k).value = v; });
|
||||
if (!$('base').value) $('base').value = 'http://10.250.50.70:4000/v1';
|
||||
if (!$('ttsUrl').value) $('ttsUrl').value = 'http://10.100.79.3:8299/tts';
|
||||
if (!$('ttsVoice').value) $('ttsVoice').value = 'baddy';
|
||||
// ext-tts default = zonos-gateway direct (streams; bypasses LiteLLM audio buffering). Migrate any stale
|
||||
// mOrpheus URL (:8299/:8210//tts) OR the LiteLLM URL (:4000) to the direct gateway, persisting so it sticks.
|
||||
if (!$('ttsUrl').value || /:(8299|8210)|\/tts(\/stream)?$|10\.250\.50\.70:4000/.test($('ttsUrl').value)){
|
||||
$('ttsUrl').value = 'http://10.100.79.3:8890/v1/audio/speech'; localStorage.setItem('gc_ttsUrl', $('ttsUrl').value);
|
||||
}
|
||||
if (!$('ttsVoice').value || $('ttsVoice').value.trim() === 'baddy'){
|
||||
$('ttsVoice').value = 'Cora'; localStorage.setItem('gc_ttsVoice', $('ttsVoice').value);
|
||||
}
|
||||
cfg.forEach(k => $(k).addEventListener('input', e => localStorage.setItem('gc_'+k, e.target.value)));
|
||||
// 🔊 auto-play toggle is a checkbox (persist .checked)
|
||||
$('ttsOn').checked = localStorage.getItem('gc_ttsOn') === '1';
|
||||
@@ -164,10 +170,14 @@ function extractQuotes(text){
|
||||
// PCM onto the shared speechHead clock; resolves when its audio is fully received (it plays on
|
||||
// while the NEXT quote starts generating). Chunk by QUOTE, not sentence — per-sentence lost prosody.
|
||||
async function streamOne(sentence, mine){
|
||||
const url = $('ttsUrl').value.trim().replace(/\/+$/,'').replace(/\/tts$/,'') + '/tts/stream';
|
||||
// ext-tts via zonos-gateway (direct, streams — bypasses LiteLLM's audio buffering): ttsUrl is the full
|
||||
// OpenAI /v1/audio/speech endpoint. The gateway relays Zonos's raw float32 PCM @ 44.1kHz — decode as
|
||||
// little-endian float32 (NOT int16) at 44100 Hz. The Bearer header is sent but the gateway ignores it.
|
||||
const url = $('ttsUrl').value.trim().replace(/\/+$/,'');
|
||||
const res = await fetch(url, {
|
||||
method:'POST', headers:{ 'Content-Type':'application/json' },
|
||||
body: JSON.stringify({ text: sentence, voice: ($('ttsVoice').value.trim() || 'baddy'), max_tokens: 2400 })
|
||||
method:'POST',
|
||||
headers:{ 'Content-Type':'application/json', 'Authorization':'Bearer ' + $('key').value },
|
||||
body: JSON.stringify({ model: 'ext-tts', input: sentence, voice: ($('ttsVoice').value.trim() || 'Cora') })
|
||||
});
|
||||
if (!res.ok || !res.body) return false;
|
||||
const reader = res.body.getReader();
|
||||
@@ -177,12 +187,12 @@ async function streamOne(sentence, mine){
|
||||
if (done || mine !== ttsGen) break; // stop if a newer reply took over
|
||||
let bytes = new Uint8Array(leftover.length + value.length);
|
||||
bytes.set(leftover); bytes.set(value, leftover.length);
|
||||
const n = bytes.length - (bytes.length % 2); // whole int16 samples only
|
||||
const n = bytes.length - (bytes.length % 4); // whole float32 samples only (ext-tts PCM)
|
||||
leftover = bytes.slice(n);
|
||||
if (!n) continue;
|
||||
const dv = new DataView(bytes.buffer, 0, n), N = n / 2, f32 = new Float32Array(N);
|
||||
for (let i = 0; i < N; i++) f32[i] = dv.getInt16(i * 2, true) / 32768;
|
||||
const abuf = audioCtx.createBuffer(1, N, 24000); abuf.getChannelData(0).set(f32);
|
||||
const dv = new DataView(bytes.buffer, 0, n), N = n / 4, f32 = new Float32Array(N);
|
||||
for (let i = 0; i < N; i++) f32[i] = dv.getFloat32(i * 4, true); // Zonos float32 PCM, little-endian
|
||||
const abuf = audioCtx.createBuffer(1, N, 44100); abuf.getChannelData(0).set(f32);
|
||||
const src = audioCtx.createBufferSource(); src.buffer = abuf; src.connect(audioCtx.destination);
|
||||
if (speechHead < audioCtx.currentTime + 0.05) speechHead = audioCtx.currentTime + 0.05; // underrun catch-up
|
||||
src.start(speechHead); speechHead += abuf.duration; // queue back-to-back on the shared clock
|
||||
@@ -193,17 +203,17 @@ async function speakQuotes(text, msg){
|
||||
if (!$('ttsOn').checked) return;
|
||||
const quotes = extractQuotes(text);
|
||||
if (!quotes.length) return;
|
||||
const chunks = quotes; // one chunk per QUOTED SECTION (keeps prosody)
|
||||
// Streaming backend (zonos-gateway) → send ALL quoted dialogue as ONE streamed request;
|
||||
// the stream itself is the incremental playback, so no per-quote chunking is needed.
|
||||
const joined = quotes.join(' ');
|
||||
const tag = document.createElement('span'); tag.textContent = ' 🔊';
|
||||
tag.style.cursor = 'pointer'; tag.title = 'mOrpheus — ' + chunks.length + ' quote(s), click to replay';
|
||||
tag.style.cursor = 'pointer'; tag.title = 'ext-tts — ' + quotes.length + ' quote(s), click to replay';
|
||||
const play = async () => {
|
||||
const mine = ++ttsGen; // supersede any in-flight playback
|
||||
primeAudio(); if (audioCtx.state === 'suspended') await audioCtx.resume();
|
||||
speechHead = audioCtx.currentTime + 0.12; // reset the shared clock for this reply
|
||||
for (const s of chunks){ // play quoted sections serially, in order
|
||||
if (mine !== ttsGen) return;
|
||||
if (!(await streamOne(s, mine))){ tag.textContent = ' 🔇'; return; }
|
||||
}
|
||||
if (mine !== ttsGen) return;
|
||||
if (!(await streamOne(joined, mine))){ tag.textContent = ' 🔇'; return; }
|
||||
if (mine === ttsGen) tag.textContent = ' 🔈';
|
||||
};
|
||||
tag.onclick = () => play().catch(() => tag.textContent = ' 🔇');
|
||||
|
||||
Reference in New Issue
Block a user