fix(asr): 过滤高置信度重复循环字幕

main/helpers/engines/builtinEngine.ts

@@ -34,6 +34,7 @@ import {
   getVadSettings,
   isReduceRepetitionEnabled,
   getNumericSetting,
+  guardAsrSubtitleCues,
 } from './transcribeShared';
 import { resolveEffectiveSettings } from './outcomePresets';
 import {
@@ -370,7 +371,14 @@ async function transcribeBuiltin(ctx: TranscribeContext): Promise<string> {
         tempAudioFile,
       );
     }
-    const formattedSrt = formatSrtContent(subtitles);
+    const guarded = guardAsrSubtitleCues(
+      subtitles,
+      formData as Record<string, unknown>,
+      settings,
+      'builtin',
+    );
+    if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+    const formattedSrt = formatSrtContent(guarded.cues);
     await fs.promises.writeFile(srtFile, formattedSrt);
 
     event.sender.send('taskFileChange', { ...file, extractSubtitle: 'done' });

main/helpers/engines/cloudAsrEngine.ts

@@ -18,7 +18,7 @@ import {
   type CloudAudioChunk,
 } from '../audioProcessor';
 import { formatSrtContent } from '../fileUtils';
-import { logMessage } from '../storeManager';
+import { logMessage, store } from '../storeManager';
 import {
   getTaskContext,
   TaskCancelledError,
@@ -35,6 +35,8 @@ import {
 import { resplitSubtitleCues } from '../subtitleSegmentation';
 import type { AsrWord } from '../../service/asr/types';
 import { getCloudProviderGate } from './cloudProviderGate';
+import { resolveEffectiveSettings } from './outcomePresets';
+import { guardAsrSubtitleCues } from './transcribeShared';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
 /** 无时间戳（text-only 模型）降级时，用更细的静音切片换取更细的粗粒度时间轴。 */
@@ -107,6 +109,10 @@ async function transcribeCloud(ctx: TranscribeContext): Promise<string> {
 
   const { tempAudioFile, srtFile } = file;
   const signal = ctx.signal ?? getTaskContext()?.signal;
+  const settings = resolveEffectiveSettings(
+    formData,
+    store.get('settings') as Record<string, unknown>,
+  );
 
   const f = formData as {
     asrProviderId?: string;
@@ -227,7 +233,14 @@ async function transcribeCloud(ctx: TranscribeContext): Promise<string> {
   throwIfSignalCancelled(signal);
   // 词级/段级路径统一补一次「裁尾」护栏（基于原始 16kHz WAV 能量）。
   const subtitles = trimSubtitleTrailingSilence(cues, tempAudioFile);
-  const formattedSrt = formatSrtContent(subtitles);
+  const guarded = guardAsrSubtitleCues(
+    subtitles,
+    formData as Record<string, unknown>,
+    settings,
+    'cloud',
+  );
+  if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+  const formattedSrt = formatSrtContent(guarded.cues);
   await fs.promises.writeFile(srtFile, formattedSrt);
 
   event.sender.send('taskProgressChange', file, 'extractSubtitle', 100);

main/helpers/engines/fasterWhisperEngine.ts

@@ -33,6 +33,7 @@ import {
   getNumericSetting,
   getWhisperLanguage,
   getFasterWhisperAntiRepetitionParams,
+  guardAsrSubtitleCues,
 } from './transcribeShared';
 import { resolveEffectiveSettings } from './outcomePresets';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
@@ -336,7 +337,14 @@ async function transcribeFasterWhisper(
     subtitles = segments.map(subtitleCueFromSegment);
   }
   subtitles = trimSubtitleTrailingSilence(subtitles, tempAudioFile);
-  const formattedSrt = formatSrtContent(subtitles);
+  const guarded = guardAsrSubtitleCues(
+    subtitles,
+    formData as Record<string, unknown>,
+    settings,
+    'faster-whisper',
+  );
+  if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+  const formattedSrt = formatSrtContent(guarded.cues);
   await fs.promises.writeFile(srtFile, formattedSrt);
 
   event.sender.send('taskProgressChange', file, 'extractSubtitle', 100);

main/helpers/engines/fireRedEngine.ts

@@ -23,6 +23,7 @@ import {
 import { resplitSubtitleCues } from '../subtitleSegmentation';
 import { buildFireRedParams } from './fireRedParams';
 import { resolveEffectiveSettings } from './outcomePresets';
+import { guardAsrSubtitleCues } from './transcribeShared';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
 /** 在途转写 id 集合：任务级并发下可能同时存在多个（排队+执行），取消须精确到 id。 */
@@ -144,7 +145,14 @@ async function transcribeFireRed(ctx: TranscribeContext): Promise<string> {
     ),
     tempAudioFile,
   );
-  const formattedSrt = formatSrtContent(subtitles);
+  const guarded = guardAsrSubtitleCues(
+    subtitles,
+    formData as Record<string, unknown>,
+    settings,
+    'firered',
+  );
+  if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+  const formattedSrt = formatSrtContent(guarded.cues);
   await fs.promises.writeFile(srtFile, formattedSrt);
 
   event.sender.send('taskProgressChange', file, 'extractSubtitle', 100);

main/helpers/engines/funasrEngine.ts

@@ -24,6 +24,7 @@ import {
 import { resplitSubtitleCues } from '../subtitleSegmentation';
 import { buildFunasrParams } from './funasrParams';
 import { resolveEffectiveSettings } from './outcomePresets';
+import { guardAsrSubtitleCues } from './transcribeShared';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
 /** 在途转写 id 集合：任务级并发下可能同时存在多个（排队+执行），取消须精确到 id。 */
@@ -154,7 +155,14 @@ async function transcribeFunasr(ctx: TranscribeContext): Promise<string> {
     ),
     tempAudioFile,
   );
-  const formattedSrt = formatSrtContent(subtitles);
+  const guarded = guardAsrSubtitleCues(
+    subtitles,
+    formData as Record<string, unknown>,
+    settings,
+    'funasr',
+  );
+  if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+  const formattedSrt = formatSrtContent(guarded.cues);
   await fs.promises.writeFile(srtFile, formattedSrt);
 
   event.sender.send('taskProgressChange', file, 'extractSubtitle', 100);

main/helpers/engines/localCliEngine.ts

@@ -3,8 +3,11 @@ import path from 'path';
 import fs from 'fs';
 import type { EngineStatus } from '../../../types/engine';
 import { logMessage, store } from '../storeManager';
+import { formatSrtContent, secondsToSubtitleTime } from '../fileUtils';
+import { parseSubtitleCues } from '../subtitleFormats';
 import { getTaskContext, TaskCancelledError } from '../taskContext';
-import { getWhisperLanguage } from './transcribeShared';
+import { resolveEffectiveSettings } from './outcomePresets';
+import { getWhisperLanguage, guardAsrSubtitleCues } from './transcribeShared';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
 /** 在途 CLI 子进程集合：任务级并发下可能同时存在多个，取消须精确到进程。 */
@@ -39,8 +42,11 @@ function transcribeLocalCli(ctx: TranscribeContext): Promise<string> {
     sourceLanguage?: string;
   };
   const whisperModel = model?.toLowerCase();
-  const settings = store.get('settings');
-  const whisperCommand = settings?.whisperCommand;
+  const settings = resolveEffectiveSettings(
+    formData,
+    store.get('settings') as Record<string, unknown>,
+  );
+  const whisperCommand = String(settings?.whisperCommand ?? '');
   const { tempAudioFile, srtFile, directory } = file;
 
   let runShell = whisperCommand
@@ -130,6 +136,33 @@ function transcribeLocalCli(ctx: TranscribeContext): Promise<string> {
         fs.renameSync(tempSrtFile, srtFile);
       }
 
+      // 外部 CLI 直接产出 SRT：解析成共享 cue 形状后走同一模型外重复护栏。
+      // 未命中时不重写文件，保留 CLI 的原始格式；解析失败也不影响已成功的转写。
+      if (fs.existsSync(srtFile)) {
+        try {
+          const parsed = parseSubtitleCues(
+            fs.readFileSync(srtFile, 'utf-8'),
+            'srt',
+          );
+          const guarded = guardAsrSubtitleCues(
+            parsed.map((cue) => [
+              secondsToSubtitleTime(cue.startMs / 1000),
+              secondsToSubtitleTime(cue.endMs / 1000),
+              cue.text,
+            ]),
+            formData as Record<string, unknown>,
+            settings,
+            'local-cli',
+          );
+          if (guarded.diagnostic) {
+            logMessage(guarded.diagnostic, 'warning');
+            fs.writeFileSync(srtFile, formatSrtContent(guarded.cues), 'utf-8');
+          }
+        } catch (error) {
+          logMessage(`local CLI repetition guard skipped: ${error}`, 'warning');
+        }
+      }
+
       event.sender.send('taskFileChange', { ...file, extractSubtitle: 'done' });
       resolve(srtFile);
     });

main/helpers/engines/qwenEngine.ts

@@ -23,6 +23,7 @@ import {
 import { resplitSubtitleCues } from '../subtitleSegmentation';
 import { buildQwenParams } from './qwenParams';
 import { resolveEffectiveSettings } from './outcomePresets';
+import { guardAsrSubtitleCues } from './transcribeShared';
 import type { TranscribeContext, TranscriptionEngineAdapter } from './types';
 
 /** 在途转写 id 集合：任务级并发下可能同时存在多个（排队+执行），取消须精确到 id。 */
@@ -142,7 +143,14 @@ async function transcribeQwen(ctx: TranscribeContext): Promise<string> {
     ),
     tempAudioFile,
   );
-  const formattedSrt = formatSrtContent(subtitles);
+  const guarded = guardAsrSubtitleCues(
+    subtitles,
+    formData as Record<string, unknown>,
+    settings,
+    'qwen',
+  );
+  if (guarded.diagnostic) logMessage(guarded.diagnostic, 'warning');
+  const formattedSrt = formatSrtContent(guarded.cues);
   await fs.promises.writeFile(srtFile, formattedSrt);
 
   event.sender.send('taskProgressChange', file, 'extractSubtitle', 100);

main/helpers/engines/transcribeShared.ts

@@ -1,7 +1,14 @@
 /**
- * 各引擎转写实现共用的纯工具：数值兜底、语言归一、SRT 时间格式化、VAD 设置归一。
- * 不依赖任何引擎实现，供 builtin / faster-whisper / localCli 适配器复用。
+ * 各引擎转写实现共用的纯工具：数值兜底、语言归一、SRT 时间格式化、VAD 设置归一、
+ * 模型外重复护栏。不依赖任何具体引擎实现，供所有 ASR 适配器复用。
  */
+import {
+  applyAsrRepetitionGuard,
+  formatRepetitionGuardDiagnostic,
+  getAsrRepetitionGuardMode,
+  type AsrSubtitleCue,
+  type RepetitionGuardStats,
+} from '../asrRepetitionGuard';
 
 export function getNumericSetting(
   value: unknown,
@@ -72,6 +79,32 @@ export function getFasterWhisperAntiRepetitionParams(
   };
 }
 
+export interface GuardedAsrSubtitleCues {
+  cues: AsrSubtitleCue[];
+  stats: RepetitionGuardStats;
+  diagnostic: string | null;
+}
+
+/**
+ * 各引擎写 SRT 前的统一出口：根据任务意图启用模型外重复护栏，并生成不含用户文本的诊断。
+ */
+export function guardAsrSubtitleCues(
+  cues: AsrSubtitleCue[],
+  formData: Record<string, unknown> | undefined,
+  effectiveSettings: Record<string, unknown> | undefined,
+  engine: string,
+): GuardedAsrSubtitleCues {
+  const taskMode = getAsrRepetitionGuardMode(formData, effectiveSettings);
+  const result = applyAsrRepetitionGuard(cues, {
+    enabled: taskMode !== 'off',
+    mode: taskMode === 'off' ? 'standard' : taskMode,
+  });
+  return {
+    ...result,
+    diagnostic: formatRepetitionGuardDiagnostic(engine, result.stats),
+  };
+}
+
 /** 从 store 的 settings 归一化出 VAD 参数（各引擎再映射到自己的字段名）。 */
 export function getVadSettings(settings: Record<string, unknown>): VadSettings {
   return {

package.json

@@ -22,6 +22,7 @@
     "native:fetch": "node scripts/fetch-sherpa-native.mjs && node scripts/fetch-whisper-addon.mjs",
     "size:check": "node scripts/check-bundle-size.mjs",
     "test:engines": "tsc scripts/test-engine-units.ts --outDir node_modules/.cache/engine-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/engine-tests/scripts/test-engine-units.js",
+    "test:asr-repetition-guard": "tsc scripts/test-asr-repetition-guard.ts --outDir node_modules/.cache/asr-repetition-guard-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck && node node_modules/.cache/asr-repetition-guard-tests/scripts/test-asr-repetition-guard.js",
     "test:ass-builder": "tsc scripts/test-ass-builder.ts --outDir node_modules/.cache/ass-builder-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/ass-builder-tests/scripts/test-ass-builder.js",
     "test:compose": "tsc scripts/compose/test-compose-builder.ts --outDir node_modules/.cache/compose-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/compose-tests/scripts/compose/test-compose-builder.js",
     "smoke:compose": "tsc scripts/compose/smoke-compose.ts --outDir node_modules/.cache/compose-tests --module commonjs --moduleResolution node --target es2019 --esModuleInterop --skipLibCheck --resolveJsonModule && node node_modules/.cache/compose-tests/scripts/compose/smoke-compose.js",

renderer/public/locales/en/tasks.json

@@ -144,7 +144,7 @@
       "title": "Custom",
       "desc": "Manually tune context length, VAD, repetition reduction and other low-level parameters.",
       "sectionTitle": "Custom parameters",
-      "sherpaNote": "For the current engine (FunASR / Qwen / FireRed), VAD is always on and there is no context-length / repetition concept. To fine-tune VAD sensitivity, use the Settings page."
+      "sherpaNote": "For the current engine (FunASR / Qwen / FireRed), VAD is always on and there are no decoder-level context or repetition controls. Cleanest & most stable enables the shared high-confidence repetition post-processor. To fine-tune VAD sensitivity, use the Settings page."
     },
     "compare": {
       "toggle": "See how the effects differ",
@@ -169,7 +169,7 @@
     "label": "Reduce repetition / hallucinated lines",
     "on": "On: drops previous-text conditioning and suppresses repeats, greatly reducing looped lines over silence/repetition; coherence on clean speech may drop slightly.",
     "off": "Off: engine defaults (keeps cross-segment coherence); long videos or silence may produce repeated or hallucinated lines.",
-    "hint": "Turn on when subtitles loop or repeat heavily. Global setting; applies to both faster-whisper and whisper.cpp."
+    "hint": "Repetition post-processing runs only with Cleanest & most stable or when this custom option is enabled. In addition to engine-native controls, it blocks sliding, exact-repeat and multi-line window cycles."
   },
   "subtitleLength": {
     "label": "Subtitle line breaking",

renderer/public/locales/zh/tasks.json

@@ -144,7 +144,7 @@
       "title": "自定义",
       "desc": "手动调整上下文长度、VAD、抗重复等底层参数。",
       "sectionTitle": "自定义参数",
-      "sherpaNote": "当前引擎（FunASR / Qwen / FireRed）的 VAD 为结构性常开，且无「上下文长度 / 抗重复」概念。如需细调 VAD 灵敏度，请到「设置」页调整。"
+      "sherpaNote": "当前引擎（FunASR / Qwen / FireRed）的 VAD 为结构性常开，且无解码层「上下文长度 / 抗重复」参数；选择「最干净最稳」会启用共享的高置信度重复后处理。如需细调 VAD 灵敏度，请到「设置」页调整。"
     },
     "compare": {
       "toggle": "看看不同效果的字幕差异",
@@ -169,7 +169,7 @@
     "label": "减少重复 / 幻觉字幕",
     "on": "已开启：断开上文条件并抑制重复，显著减少静音/重复段的循环字幕；干净语音的连贯性可能略降。",
     "off": "已关闭：使用引擎默认（保留上文连贯性）；长视频或静音处可能出现重复或幻觉字幕。",
-    "hint": "遇到字幕大段重复/鬼畜时开启。全局设置，对 faster-whisper 与 whisper.cpp 同时生效。"
+    "hint": "仅在选择「最干净最稳」或自定义开启本项时运行重复后处理；除引擎自身的抗重复参数外，还会拦截滑动、完全重复与多条字幕窗口循环。"
   },
   "subtitleLength": {
     "label": "字幕断句方式",