Web Speech API sends audio to Google's servers for transcription. For privacy-sensitive workflows — especially when narrating proprietary UI or internal tools — we need a local alternative that keeps audio on-device.
Options investigated:
- Whisper.cpp via WebAssembly — community WASM builds exist. Would run entirely in the extension's offscreen document or a worker. Latency is higher than streaming Web Speech API but acceptable for our use case (short narration clips, not real-time captions).
- Chrome's on-device speech recognition — behind
chrome://flags/#enable-on-device-speech-recognition. Not yet stable or available on all platforms, but worth tracking.
Implementation notes:
- The transcription interface in the sidepanel already receives text chunks — swapping the backend should be transparent to the rest of the pipeline.
- Model download size (~40MB for tiny.en) needs a first-run UX.
- This would make PointDev fully offline-capable when combined with the existing local-only capture pipeline.
Related to NLnet milestone M4 (privacy and offline support).
Web Speech API sends audio to Google's servers for transcription. For privacy-sensitive workflows — especially when narrating proprietary UI or internal tools — we need a local alternative that keeps audio on-device.
Options investigated:
chrome://flags/#enable-on-device-speech-recognition. Not yet stable or available on all platforms, but worth tracking.Implementation notes:
Related to NLnet milestone M4 (privacy and offline support).