fix(asr): 修复 Windows 本地长语音只识别前段

[Duyi-Wang]

User description

摘要

修复 Windows 本地 ASR 在长语音场景下可能只识别前一段的问题。

主要原因是 foundry-local-whisper 使用 Whisper 系模型，单次解码窗口约 30 秒；当录音超过这个窗口时，单次把整段 WAV 交给 Foundry SDK 可能只返回前段结果。这个 PR 将 Foundry Local Whisper 的 PCM 输入按 30 秒分片转写，再合并分片文本；同时把 Windows 本地 ASR 的转写超时从固定 30 秒改为按音频时长动态增长。

未关联 issue。

修复 / 新增 / 改进

• foundry-local-whisper 录音结束后按 30 秒 PCM chunk 分片，逐段生成临时 WAV 并调用 Foundry SDK。
• 复用 Whisper batch ASR 的 chunk split / transcript join 逻辑，避免中英文、CJK 标点拼接时产生明显断裂。
• 为 Foundry / Sherpa provider 增加 buffer_duration_ms()，coordinator 可按实际录音长度计算 timeout。
• 普通听写、QA、overlay QA、历史重新转录路径都接入 Windows 本地 ASR 动态 timeout。
• 将 Windows Foundry / Sherpa 本地 ASR timeout 调整为 max(30s, ceil(audio_s) + 20s)，避免长音频撞固定 30 秒超时。
• 增加 Foundry 分片、buffer duration、Sherpa buffer duration、Windows 本地 ASR timeout 相关测试。

兼容

• 不包含：

  • 云端 ASR 协议变更。
  • 录音采集链路变更。
  • 插入 / polish 流程变更。
  • 用户配置迁移或模型配置变更。

• 对现有用户 / 本地环境 / 构建流程的影响：

  • 短语音仍使用 30 秒 timeout 保底，行为基本不变。
  • 长语音在 foundry-local-whisper 下会拆成多段本地识别，整体耗时会随音频长度增加，但可以覆盖完整录音。
  • sherpa-onnx-local 仅调整 timeout 预算，不改变识别输入和模型调用方式。

测试计划

• 命令：cargo test

• 结果：本地 Rust 单测通过，546 passed; 0 failed

• 证据路径：本地终端运行结果

• 命令：fork 上的 GitHub Actions CI

• 结果：通过

• 证据路径：https://github.com/Duyi-Wang/openless/actions/runs/28922670487

• 命令：fork 上的桌面安装包构建 Release Tauri (cross-platform)

• 结果：通过

• 证据路径：https://github.com/Duyi-Wang/openless/actions/runs/28922670561

• 命令：fork 上的 Android APK 构建

• 结果：通过

• 证据路径：https://github.com/Duyi-Wang/openless/actions/runs/28922670488

---

PR Type

Bug fix

---

Description

• Split Foundry Local Whisper PCM into 30s chunks to prevent truncation of long recordings

• Added dynamic transcribe timeout (max(30s, ceil(audio_s)+20s)) for Windows local ASR

• Exposed buffer_duration_ms() from Foundry and Sherpa providers

• Applied dynamic timeout to dictation, QA, overlay QA, and re-transcription paths

---

Diagram Walkthrough

flowchart LR
  audio["Recorded PCM (long)"] --> split["Split into 30s chunks"]
  split --> chunk1["Chunk 1"] & chunk2["Chunk 2"] & chunkN["Chunk N"]
  chunk1 --> transcribe1["Transcribe with remaining timeout"]
  chunk2 --> transcribe2["Transcribe with remaining timeout"]
  chunkN --> transcribeN["Transcribe with remaining timeout"]
  transcribe1 & transcribe2 & transcribeN --> join["Join transcripts"]
  join --> result["Full transcript"]
  audio -- "buffer_duration_ms()" --> timeout["Dynamic timeout = max(30s, ceil(audio_s)+20s)"]
  timeout --> transcribe1 & transcribe2 & transcribeN

Loading

File Walkthrough

Relevant files

Enhancement

foundry_provider.rs Chunk Foundry Whisper PCM into 30s segments                            openless-all/app/src-tauri/src/asr/local/foundry_provider.rs Added FOUNDRY_WHISPER_CHUNK_LIMIT_MS constant (30s) and buffer_duration_ms() method Modified transcribe() to split PCM into 30s chunks and transcribe each with remaining deadline timeout Added helper remaining_transcribe_timeout() and tests for chunking, timeout calculation, and buffer reporting +111/-17 sherpa_provider.rs Add buffer_duration_ms to Sherpa ONNX provider                      openless-all/app/src-tauri/src/asr/local/sherpa_provider.rs Added buffer_duration_ms() for both Offline (PCM buffer) and Online (worker audio bytes) modes Added test to verify buffer duration reported without consuming buffer +25/-0    whisper.rs Export PCM split and transcript join functions                      openless-all/app/src-tauri/src/asr/whisper.rs Made split_pcm_by_duration() and join_transcript_chunks() pub(crate) for reuse by Foundry provider +2/-2      coordinator.rs Replace fixed timeouts with dynamic per-audio timeout        openless-all/app/src-tauri/src/coordinator.rs Unified fixed timeout functions into dynamic windows_local_asr_transcribe_timeout(audio_secs) Updated Foundry and Sherpa transcribe calls in re_transcribe_current_session to use audio-relative timeout Updated tests: short audio floors at 30s, long audio scales (e.g., 65s → 85s) +32/-23  dictation.rs Use dynamic timeout in dictation end session                          openless-all/app/src-tauri/src/coordinator/dictation.rs Updated end_session for Foundry Local Whisper and Sherpa ONNX to compute timeout from buffer_duration_ms() Added logging of audio duration and timeout +16/-8    qa_session.rs Use dynamic timeout in QA transcription paths                        openless-all/app/src-tauri/src/coordinator/qa_session.rs Updated transcribe_overlay_dictation_asr and end_qa_session for Foundry and Sherpa to compute dynamic timeout Added logging in QA transcription paths +22/-16

---

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 3f2c6a4)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ No major issues detected

[github-actions]

Persistent review updated to latest commit 3f2c6a4