fix(qwen): 去掉字幕里的 language Chinese / asr_text 模板残留

extraResources/sherpa/README.md

@@ -3,8 +3,10 @@
 封装 JS 复制自 npm `sherpa-onnx-node@1.13.2`（Apache-2.0）。
 
 - `vendor/addon.js` 已替换为自定义加载器：从环境变量 `SHERPA_ONNX_LIB_DIR`
-  用 `process.dlopen` 加载 `sherpa-onnx.node`（其余 `vendor/*.js` 原样保留，
-  它们都经 `require('./addon.js')` 取原生模块）。
+  用 `process.dlopen` 加载 `sherpa-onnx.node`；vendor 文件都经
+  `require('./addon.js')` 取原生模块。
+- `vendor/json-result.js` 由 `non-streaming-asr.js` 使用，处理原生识别结果中
+  的未转义控制字符；其余 `vendor/*.js` 原样保留。
 - `vendor/addon-static-import.js` 在自定义加载器下不再被引用（保留以便升级对照）。
 
 原生库**不在此处**，按需下载到 `userData/sherpa-onnx/current/`

extraResources/sherpa/vendor/json-result.js

@@ -0,0 +1,74 @@
+'use strict';
+
+// Native sherpa results are JSON strings. Some Qwen outputs may contain raw
+// C0 control characters inside the transcript text, which JSON.parse rejects.
+function sanitizeJsonControlCharacters(jsonStr) {
+  let sanitized = '';
+  let inString = false;
+  let escaped = false;
+
+  for (const char of String(jsonStr)) {
+    const code = char.charCodeAt(0);
+
+    if (inString) {
+      if (escaped) {
+        sanitized += char;
+        escaped = false;
+        continue;
+      }
+      if (char === '\\') {
+        sanitized += char;
+        escaped = true;
+        continue;
+      }
+      if (char === '"') {
+        sanitized += char;
+        inString = false;
+        continue;
+      }
+      if (code <= 0x1f) {
+        sanitized += escapeControlCharacter(char);
+        continue;
+      }
+      sanitized += char;
+      continue;
+    }
+
+    if (char === '"') {
+      inString = true;
+      sanitized += char;
+    } else if (code > 0x1f || char === '\t' || char === '\n' || char === '\r') {
+      // Tabs, line feeds, and carriage returns are valid JSON whitespace
+      // outside strings. Other C0 controls are never valid JSON tokens.
+      sanitized += char;
+    }
+  }
+
+  return sanitized;
+}
+
+function escapeControlCharacter(char) {
+  switch (char) {
+    case '\b':
+      return '\\b';
+    case '\f':
+      return '\\f';
+    case '\n':
+      return '\\n';
+    case '\r':
+      return '\\r';
+    case '\t':
+      return '\\t';
+    default:
+      return `\\u${char.charCodeAt(0).toString(16).padStart(4, '0')}`;
+  }
+}
+
+function parseJsonResult(jsonStr) {
+  return JSON.parse(sanitizeJsonControlCharacters(jsonStr));
+}
+
+module.exports = {
+  parseJsonResult,
+  sanitizeJsonControlCharacters,
+};

extraResources/sherpa/vendor/non-streaming-asr.js

@@ -8,6 +8,7 @@
  */
 
 const addon = require('./addon.js');
+const { parseJsonResult } = require('./json-result.js');
 
 /**
  * Internal symbol to mark async-created recognizers.
@@ -127,7 +128,7 @@ class OfflineRecognizer {
       this.handle,
       stream.handle,
     );
-    return JSON.parse(jsonStr);
+    return parseJsonResult(jsonStr);
   }
 
   /**
@@ -137,7 +138,7 @@ class OfflineRecognizer {
    */
   getResult(stream) {
     const jsonStr = addon.getOfflineStreamResultAsJson(stream.handle);
-    return JSON.parse(jsonStr);
+    return parseJsonResult(jsonStr);
   }
 }
 

extraResources/sherpa/worker/sherpa-worker.js

@@ -243,6 +243,104 @@ function ensureLoaded(req) {
   }
 }
 
+// Qwen3-ASR 聊天模板：language Chinese<asr_text>正文
+// 逻辑与 main/helpers/engines/qwenText.ts 保持一致。
+const QWEN_TEMPLATE_LEFTOVER =
+  /^(?:language|chinese|english|none|zh|en|auto)$/i;
+
+const QWEN_LATIN_KEEP = new Set([
+  'ok',
+  'okay',
+  'yes',
+  'no',
+  'yeah',
+  'yep',
+  'nope',
+  'right',
+  'so',
+  'well',
+  'um',
+  'uh',
+  'ah',
+  'oh',
+  'wow',
+  'next',
+  'thanks',
+  'hello',
+  'hi',
+  'bye',
+  'alpha',
+  'beta',
+  'gamma',
+  'delta',
+  'epsilon',
+  'theta',
+  'lambda',
+  'mu',
+  'nu',
+  'xi',
+  'pi',
+  'rho',
+  'sigma',
+  'tau',
+  'phi',
+  'psi',
+  'chi',
+  'eta',
+  'omega',
+]);
+
+function stripQwenCueDecorations(text) {
+  return text
+    .replace(/^[("'“‘]+/, '')
+    .replace(/[)"'”’.,!?。！？]+$/, '')
+    .trim();
+}
+
+function isQwenFormulaToken(token) {
+  return (
+    /^[A-Za-z]$/.test(token) ||
+    /^[A-Za-z]\d{1,3}$/.test(token) ||
+    /^[A-Z]{2,4}$/.test(token) ||
+    /^[A-Z]{2,4}\d{1,3}$/.test(token)
+  );
+}
+
+function isQwenHallucinatedCue(text) {
+  const token = stripQwenCueDecorations(text);
+  if (!token) return true;
+  if (QWEN_TEMPLATE_LEFTOVER.test(token)) return true;
+  if (/[\u4e00-\u9fff]/.test(text)) return false;
+  if (/\s/.test(token)) return false;
+  if (/^[\d.]+$/.test(token)) return false;
+  if (isQwenFormulaToken(token)) return false;
+  if (QWEN_LATIN_KEEP.has(token.toLowerCase())) return false;
+  return /^[A-Za-z][A-Za-z0-9]{1,23}$/.test(token);
+}
+
+function sanitizeQwenAsrText(raw) {
+  if (typeof raw !== 'string' || raw.length === 0) return '';
+  let text = raw;
+  if (/<\s*asr_text\s*>/i.test(text)) {
+    const parts = text.split(/<\s*asr_text\s*>/i);
+    text = parts[parts.length - 1] || '';
+  }
+  text = text.replace(/<\s*\/\s*asr_text\s*>/gi, '');
+  text = text
+    .replace(/^\*+\s*/, '')
+    .replace(/\s*\*+$/, '')
+    .trim();
+  if (
+    /^language\s*[:：]?\s*[A-Z][A-Za-z]+(?:[\s-][A-Z][A-Za-z]+){0,2}$/.test(
+      text,
+    )
+  ) {
+    return '';
+  }
+  if (isQwenHallucinatedCue(text)) return '';
+  return text;
+}
+
 function postCancelled(id) {
   cancelled.delete(id);
   channel.post({
@@ -273,7 +371,9 @@ async function transcribe(req) {
       const r = await recognizer.decodeAsync(stream);
       const start = seg.start / SAMPLE_RATE;
       const end = (seg.start + seg.samples.length) / SAMPLE_RATE;
-      const text = r && r.text ? r.text.trim() : '';
+      const rawText = r && r.text ? r.text.trim() : '';
+      const text =
+        req.modelType === 'qwen3_asr' ? sanitizeQwenAsrText(rawText) : rawText;
       if (text) segments.push({ start, end, text });
     }
   };

main/helpers/engines/qwenEngine.ts

@@ -22,6 +22,7 @@ import {
 } from '../subtitleTiming';
 import { resplitSubtitleCues } from '../subtitleSegmentation';
 import { buildQwenParams } from './qwenParams';
+import { sanitizeQwenAsrText } from './qwenText';
 import { resolveEffectiveSettings } from './outcomePresets';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
@@ -137,7 +138,13 @@ async function transcribeQwen(ctx: TranscribeContext): Promise<string> {
 
   const subtitles = trimSubtitleTrailingSilence(
     resplitSubtitleCues(
-      (transcription?.segments || []).map(subtitleCueFromSegment),
+      (transcription?.segments || [])
+        .map((segment) => ({
+          ...segment,
+          text: sanitizeQwenAsrText(segment.text),
+        }))
+        .filter((segment) => Boolean(segment.text))
+        .map(subtitleCueFromSegment),
       formData as Record<string, unknown>,
     ),
     tempAudioFile,

main/helpers/engines/qwenText.ts

@@ -0,0 +1,111 @@
+/**
+ * Qwen3-ASR 解码原文常带聊天模板前缀，例如：
+ *   language Chinese<asr_text>所以
+ *   **language Chinese<asr_text>开玩笑 **
+ * 剥掉语种头、asr_text 标签和装饰星号，只留听写正文。
+ *
+ * 静音/幻听时还会整条吐出拉丁碎片（demand、imageUrl、Finds.）。
+ * 第一性：没有汉字的单 token 默认不是中文课听写；
+ * 只保留公式字母、数学符号名、口语短回应和正常英文句子。
+ */
+
+const QWEN_TEMPLATE_LEFTOVER =
+  /^(?:language|chinese|english|none|zh|en|auto)$/i;
+
+/** 中文 STEM 课里单独出现也合理的拉丁短词。 */
+const QWEN_LATIN_KEEP = new Set([
+  'ok',
+  'okay',
+  'yes',
+  'no',
+  'yeah',
+  'yep',
+  'nope',
+  'right',
+  'so',
+  'well',
+  'um',
+  'uh',
+  'ah',
+  'oh',
+  'wow',
+  'next',
+  'thanks',
+  'hello',
+  'hi',
+  'bye',
+  'alpha',
+  'beta',
+  'gamma',
+  'delta',
+  'epsilon',
+  'theta',
+  'lambda',
+  'mu',
+  'nu',
+  'xi',
+  'pi',
+  'rho',
+  'sigma',
+  'tau',
+  'phi',
+  'psi',
+  'chi',
+  'eta',
+  'omega',
+]);
+
+function stripCueDecorations(text: string): string {
+  return text
+    .replace(/^[("'“‘]+/, '')
+    .replace(/[)"'”’.,!?。！？]+$/, '')
+    .trim();
+}
+
+function isFormulaToken(token: string): boolean {
+  return (
+    /^[A-Za-z]$/.test(token) ||
+    /^[A-Za-z]\d{1,3}$/.test(token) ||
+    /^[A-Z]{2,4}$/.test(token) ||
+    /^[A-Z]{2,4}\d{1,3}$/.test(token)
+  );
+}
+
+function isQwenHallucinatedCue(text: string): boolean {
+  const token = stripCueDecorations(text);
+  if (!token) return true;
+  if (QWEN_TEMPLATE_LEFTOVER.test(token)) return true;
+  if (/[\u4e00-\u9fff]/.test(text)) return false;
+  if (/\s/.test(token)) return false;
+  if (/^[\d.]+$/.test(token)) return false;
+  if (isFormulaToken(token)) return false;
+  if (QWEN_LATIN_KEEP.has(token.toLowerCase())) return false;
+  return /^[A-Za-z][A-Za-z0-9]{1,23}$/.test(token);
+}
+
+export function sanitizeQwenAsrText(raw: string | null | undefined): string {
+  if (typeof raw !== 'string' || raw.length === 0) return '';
+
+  let text = raw;
+  if (/<\s*asr_text\s*>/i.test(text)) {
+    const parts = text.split(/<\s*asr_text\s*>/i);
+    text = parts[parts.length - 1] ?? '';
+  }
+
+  text = text.replace(/<\s*\/\s*asr_text\s*>/gi, '');
+  text = text
+    .replace(/^\*+\s*/, '')
+    .replace(/\s*\*+$/, '')
+    .trim();
+
+  if (
+    /^language\s*[:：]?\s*[A-Z][A-Za-z]+(?:[\s-][A-Z][A-Za-z]+){0,2}$/.test(
+      text,
+    )
+  ) {
+    return '';
+  }
+
+  if (isQwenHallucinatedCue(text)) return '';
+  return text;
+}

package.json

@@ -38,6 +38,7 @@
     "test:custom-languages": "tsc scripts/test-custom-languages.ts --outDir node_modules/.cache/custom-language-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck && node node_modules/.cache/custom-language-tests/scripts/test-custom-languages.js",
     "test:refine": "tsc scripts/test-refine-units.ts --outDir node_modules/.cache/refine-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/refine-tests/scripts/test-refine-units.js",
     "test:speaker-diarization": "node scripts/test-speaker-diarization-config.cjs && tsc scripts/test-speaker-diarization.ts --outDir node_modules/.cache/speaker-diarization-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck && node node_modules/.cache/speaker-diarization-tests/scripts/test-speaker-diarization.js",
+    "test:sherpa-json": "node scripts/test-sherpa-json-result.cjs",
     "test:proofread-speakers": "tsc scripts/test-proofread-speakers.ts --outDir node_modules/.cache/proofread-speaker-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck && node node_modules/.cache/proofread-speaker-tests/scripts/test-proofread-speakers.js",
     "test:manuscript": "tsc scripts/test-manuscript-matching.ts --outDir node_modules/.cache/manuscript-tests --module commonjs --moduleResolution node --target es2022 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/manuscript-tests/scripts/test-manuscript-matching.js",
     "longgap:gen": "tsc scripts/longgap/gen-audio.ts --outDir node_modules/.cache/longgap --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/longgap/scripts/longgap/gen-audio.js",