vh/esh-pfi-infrastructure
1
0
Fork 0
Code Issues 7 Pull Requests Actions Packages Projects Releases Wiki Activity

fix(gateway-chat): chunk by quoted section, not sentence (prosody)

Browse Source 
Per-sentence chunking generated each sentence cold, flattening intonation/prosody that
spans the whole quoted line. Chunk by QUOTED SECTION instead — each contiguous quote is
generated whole (max_tokens 2400) so its prosody stays intact; multiple quotes in a reply
still play serially on the shared clock. extractQuotes already returns exactly these spans;
dropped splitSentences.
This commit is contained in:
Vuong Hoang
2026-07-09 01:54:39 -07:00
parent a1f3023f70
commit f295cc1f46
2 changed files with 14 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
stacks/gateway-chat/conf/index.html
+7 -9
View File
@@ -160,16 +160,14 @@ function extractQuotes(text){
while ((m = re.exec(text)) !== null){ const q = (m[1] || m[2] || '').trim(); if (q) out.push(q); }
return out;
}
function splitSentences(t){
return (t.match(/[^.!?…]+[.!?…]+["”’')\]]*\s*|[^.!?…]+$/g) || [t]).map(s => s.trim()).filter(Boolean);
}
// stream ONE sentence and queue its PCM onto the shared speechHead clock; resolves when the
// sentence's audio is fully received (it plays on while the NEXT sentence starts generating).
// stream ONE quoted section (whole, so intonation/prosody across it is preserved) and queue its
// 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$/,'') + '/tts/stream';
const res = await fetch(url, {
method:'POST', headers:{ 'Content-Type':'application/json' },
body: JSON.stringify({ text: sentence, voice: ($('ttsVoice').value.trim() || 'baddy'), max_tokens: 900 })
body: JSON.stringify({ text: sentence, voice: ($('ttsVoice').value.trim() || 'baddy'), max_tokens: 2400 })
});
if (!res.ok || !res.body) return false;
const reader = res.body.getReader();
@@ -195,14 +193,14 @@ async function speakQuotes(text, msg){
if (!$('ttsOn').checked) return;
const quotes = extractQuotes(text);
if (!quotes.length) return;
const sentences = splitSentences(quotes.join(' ')); // pre-chunk quoted text by sentence
const chunks = quotes; // one chunk per QUOTED SECTION (keeps prosody)
const tag = document.createElement('span'); tag.textContent = ' 🔊';
tag.style.cursor = 'pointer'; tag.title = 'mOrpheus — ' + sentences.length + ' sentence(s), click to replay';
tag.style.cursor = 'pointer'; tag.title = 'mOrpheus — ' + chunks.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 sentences){ // play sentences serially, in order
for (const s of chunks){ // play quoted sections serially, in order
if (mine !== ttsGen) return;
if (!(await streamOne(s, mine))){ tag.textContent = ' 🔇'; return; }
}
tools/gateway-chat.html
+7 -9
View File
@@ -160,16 +160,14 @@ function extractQuotes(text){
while ((m = re.exec(text)) !== null){ const q = (m[1] || m[2] || '').trim(); if (q) out.push(q); }
return out;
}
function splitSentences(t){
return (t.match(/[^.!?…]+[.!?…]+["”’')\]]*\s*|[^.!?…]+$/g) || [t]).map(s => s.trim()).filter(Boolean);
}
// stream ONE sentence and queue its PCM onto the shared speechHead clock; resolves when the
// sentence's audio is fully received (it plays on while the NEXT sentence starts generating).
// stream ONE quoted section (whole, so intonation/prosody across it is preserved) and queue its
// 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$/,'') + '/tts/stream';
const res = await fetch(url, {
method:'POST', headers:{ 'Content-Type':'application/json' },
body: JSON.stringify({ text: sentence, voice: ($('ttsVoice').value.trim() || 'baddy'), max_tokens: 900 })
body: JSON.stringify({ text: sentence, voice: ($('ttsVoice').value.trim() || 'baddy'), max_tokens: 2400 })
});
if (!res.ok || !res.body) return false;
const reader = res.body.getReader();
@@ -195,14 +193,14 @@ async function speakQuotes(text, msg){
if (!$('ttsOn').checked) return;
const quotes = extractQuotes(text);
if (!quotes.length) return;
const sentences = splitSentences(quotes.join(' ')); // pre-chunk quoted text by sentence
const chunks = quotes; // one chunk per QUOTED SECTION (keeps prosody)
const tag = document.createElement('span'); tag.textContent = ' 🔊';
tag.style.cursor = 'pointer'; tag.title = 'mOrpheus — ' + sentences.length + ' sentence(s), click to replay';
tag.style.cursor = 'pointer'; tag.title = 'mOrpheus — ' + chunks.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 sentences){ // play sentences serially, in order
for (const s of chunks){ // play quoted sections serially, in order
if (mine !== ttsGen) return;
if (!(await streamOne(s, mine))){ tag.textContent = ' 🔇'; return; }
}