@olares/lares-core 0.15.26 → 0.15.27

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@olares/lares-core",
-  "version": "0.15.26",
+  "version": "0.15.27",
   "description": "Lares 共享业务核心：Router、workspace、文件/媒体、Olares 身份（无 UI、无平台壳）",
   "type": "module",
   "repository": {

package/router/tts.js

@@ -1,3 +1,4 @@
+import { speechInput } from "../voice/tts-input.js";
 import { fetchTtsVoices } from "../voice/tts-voice.js";
 import { VoiceError } from "./stt.js";
 import { routerShimBaseUrl } from "./gateway.js";
@@ -54,8 +55,9 @@ export function pcmToWav(pcm, sampleRate = PCM_RATE) {
  * @returns {Promise<Buffer>}
  */
 export async function synthesize(request) {
-  const text = request.text.trim();
-  if (!text) throw new VoiceError("voice_config_invalid", 400, "nothing to speak");
+  const spoken = request.text.trim();
+  if (!spoken) throw new VoiceError("voice_config_invalid", 400, "nothing to speak");
+  const text = speechInput(request.model, spoken);
   let res;
   try {
     res = await fetch(`${routerShimBaseUrl()}/audio/speech`, {
@@ -92,8 +94,9 @@ export async function synthesize(request) {
  * @param {(chunk: Buffer) => Promise<void> | void} onChunk
  */
 export async function synthesizeStream(request, onChunk) {
-  const text = request.text.trim();
-  if (!text) throw new VoiceError("voice_config_invalid", 400, "nothing to speak");
+  const spoken = request.text.trim();
+  if (!spoken) throw new VoiceError("voice_config_invalid", 400, "nothing to speak");
+  const text = speechInput(request.model, spoken);
   let res;
   try {
     const timeout = AbortSignal.timeout(REQUEST_TIMEOUT_MS);

package/voice/speech-segmenter.js

@@ -20,30 +20,49 @@ function boundaries(text) {
   return { positions, balanced: !stack.length };
 }
 
-function sentences(text, closed) {
-  const { positions, balanced } = boundaries(text);
-  const ends = new Set();
+function cjkCuts(text, positions) {
+  const cuts = [];
+  for (let i = 0; i < text.length; i++) {
+    if (!/[。！？]/.test(text[i])) continue;
+    let end = i + 1;
+    while (/[”’」』）)\]]/.test(text[end] || "") && end < text.length) end++;
+    if (positions.has(end)) cuts.push(end);
+  }
+  return [...new Set(cuts)].sort((a, b) => a - b);
+}
+
+// sentencex is an English segmenter. A Chinese 。！？ already closed the previous
+// sentence; feeding that terminator back in makes it treat "Mr." as another end.
+function sentenceCuts(slice, sliceClosed) {
+  const parts = segment("en", slice);
+  const ends = [];
   let from = 0;
-  const parts = segment("en", text);
   for (let i = 0; i < parts.length; i++) {
-    const at = text.indexOf(parts[i], from); if (at < 0) continue;
+    const at = slice.indexOf(parts[i], from);
+    if (at < 0) continue;
     from = at + parts[i].length;
-    if (i < parts.length - 1 || closed) ends.add(from);
+    if (i < parts.length - 1) ends.push(from);
+    else if (sliceClosed) ends.push(slice.length);
   }
-  for (let i = 0; i < text.length; i++) {
-    if (/[。！？]/.test(text[i])) {
-      let end = i + 1;
-      while (/[”’」』）)\]]/.test(text[end] || "") && end < text.length) end++;
-      ends.add(end);
-    }
+  return ends;
+}
+
+function sentences(text, closed) {
+  const { positions, balanced } = boundaries(text);
+  const ends = new Set();
+  const cuts = cjkCuts(text, positions);
+  let start = 0;
+  for (const cut of cuts) {
+    for (const end of sentenceCuts(text.slice(start, cut), true)) ends.add(start + end);
+    start = cut;
   }
-  if (closed) ends.add(text.length);
+  for (const end of sentenceCuts(text.slice(start), closed)) ends.add(start + end);
   const out = []; let consumed = 0;
   for (const end of [...ends].sort((a, b) => a - b)) {
     if (!positions.has(end) && !(closed && end === text.length)) continue;
     const sentence = text.slice(consumed, end).trim();
     if (sentence.includes("\uFFFC")) break;
-    if (sentence) out.push(sentence);
+    if (/[\p{L}\p{N}]/u.test(sentence)) out.push(sentence);
     consumed = end;
   }
   return { out, consumed, balanced };

package/voice/stream-client.js

@@ -49,7 +49,7 @@ export async function playAudioStream(options) {
     if (!progressNoted) { progressNoted = true; streamMetric(options, { stage: "playback_progress_sent", frames }); }
   },
     onMetric: (event) => { if (event.stage === "playback_started") options.onStatus?.("playing"); streamMetric(options, { ...event, durationMs: Date.now() - started }); } });
-  let chain = Promise.resolve(), complete = false, firstReceived = false, ready = false, paused = false, segmentIndex = -1;
+  let chain = Promise.resolve(), complete = false, ending = false, firstReceived = false, ready = false, paused = false, segmentIndex = -1;
   options.controls?.({
     get canPause() { return ready; },
     pause() { if (!ready || paused) return; paused = true; player.pause(); file.pause(); if (connection.readyState === 1) connection.send(JSON.stringify({ type: "pause" })); },
@@ -69,10 +69,14 @@ export async function playAudioStream(options) {
       options.signal?.addEventListener("abort", cancelled, { once: true });
       connection.addEventListener("message", ({ data }) => {
         if (typeof data !== "string" && !firstReceived) { firstReceived = true; streamMetric(options, { stage: "playback_first_received", durationMs: Date.now() - started }); }
+        let event = null;
+        if (typeof data === "string") {
+          try { event = JSON.parse(data); } catch (error) { chain = chain.then(() => { throw error; }).catch(fail); return; }
+          if (event?.type === "done") ending = true;
+        }
         chain = chain.then(async () => {
           if (options.signal?.aborted) throw new Error("voice_cancelled");
           if (typeof data !== "string") { player.write(new Uint8Array(data)); return; }
-          const event = JSON.parse(data);
           if (event.type === "segment") {
             if (event.index !== segmentIndex + 1) throw new Error("voice_reply_sequence");
             segmentIndex = event.index;
@@ -88,7 +92,7 @@ export async function playAudioStream(options) {
       connection.addEventListener("error", () => fail(dropped()), { once: true });
       connection.addEventListener("close", () => {
         options.signal?.removeEventListener("abort", cancelled);
-        if (!complete) { const error = dropped(); noteFailure(error); reject(error); }
+        if (!complete && !ending) { const error = dropped(); noteFailure(error); reject(error); }
       }, { once: true });
       if (options.signal?.aborted) cancelled();
     });

package/voice/stream-gateway.js

@@ -159,7 +159,9 @@ export function createAudioStreamGateway({ config, models, replies, synthesize =
               else if (event.type === "error") throw fault(event.code || "voice_stream_failed");
             }
           };
-          void pump().catch((cause) => { if (!abort.signal.aborted) error(cause); }).finally(() => { unsub(); close(); });
+          // `done` means the bytes have been handed over. The browser still has up to
+          // the two-second send-ahead queued. Closing now is a disconnect and stops that tail.
+          void pump().catch((cause) => { if (!abort.signal.aborted) error(cause); }).finally(() => unsub());
         }
       });
     },

package/voice/tts-input.js

@@ -0,0 +1,10 @@
+const QWEN_12HZ = /qwen3-tts|audioqwen3tts/i;
+
+/** 12Hz Qwen drops its last codec frames. This tail is not spoken, so those frames land on it. */
+export const QWEN_SPEECH_TAIL = " ^.";
+
+export function speechInput(model, text) {
+  const body = String(text ?? "");
+  if (!QWEN_12HZ.test(String(model ?? ""))) return body;
+  return body.endsWith(QWEN_SPEECH_TAIL) ? body : `${body}${QWEN_SPEECH_TAIL}`;
+}

package/voice/tts-preview.js

@@ -1,8 +1,10 @@
+import { speechInput } from "./tts-input.js";
+
 /** Explicit audition is independent of automatic read-aloud and business sessions. */
 export async function previewTts({ model, voice, text, signal }, { fetcher = fetch, createAudio = url => new Audio(url) } = {}) {
   const response = await fetcher("/llm/v1/audio/speech", {
     method: "POST", headers: { "content-type": "application/json" }, signal,
-    body: JSON.stringify({ model, input: text, response_format: "wav", ...(voice ? { voice } : {}) }),
+    body: JSON.stringify({ model, input: speechInput(model, text), response_format: "wav", ...(voice ? { voice } : {}) }),
   });
   if (!response.ok) throw new Error("tts_preview_failed");
   const blob = await response.blob(); signal.throwIfAborted();