Skip to content

plan.md

Latest commit

 

History

History
33 lines (27 loc) · 1.87 KB

plan.md

File metadata and controls

33 lines (27 loc) · 1.87 KB

Fix Plan: All Audit Findings

Files to modify (in order)

1. extensions/voice/sherpa-engine.ts

  • P0: decode(stream) → await recognizer.decodeAsync(stream), make transcribeBuffer async
  • H1: Add language to cache key: { modelId, language, recognizer }
  • H2: Strip regional suffix: language.split("-")[0] before passing to Whisper/SenseVoice
  • P1: pcmToFloat32 use Int16Array typed view instead of readInt16LE loop
  • L7: Remove dead LD_LIBRARY_PATH/DYLD_LIBRARY_PATH env var setting (add comment explaining why)
  • H9+M20: Add early process.arch === "arm" check and musl detection in initSherpa()

2. extensions/voice/local.ts

  • H8: Add if (session.closed) return; after each await in stopLocalSession
  • H7: Wrap transcribeInProcess in timeout (120s)
  • M19: Pass AbortSignal through to ensureModelDownloaded
  • P2: Remove Buffer.from(chunk) → push chunk directly; clear audioChunks after concat
  • M2: Fix tier: whisper-medium → "standard", sensevoice-small → "edge", gigaam-v3 → "edge"

3. extensions/voice/model-download.ts

  • H4: Check resp.status === 200 when startByte > 0 → switch to overwrite mode
  • M4+L1: Add backpressure (drain event) + remove Buffer.from(value) → write value directly

4. extensions/voice/device.ts

  • M10: Add /dev/nvidia0 pre-check before spawning nvidia-smi
  • L2: Add 32-bit cgroup sentinels + heuristic (bytes > hostRamBytes → unlimited)
  • L4: Add Podman detection (/run/.containerenv)

5. extensions/voice.ts

  • M15: Guard checkLocalServer calls with if (config.localEndpoint) in /voice test
  • M17: Add clearRecognizerCache() in session_switch handler
  • M18: Add clearRecognizerCache() in /voice-language handler
  • H6: Fix Windows ffmpeg: enumerate dshow devices instead of audio=default