JS speech-to-text exists in three stacks; Chrome Web Speech is not local, WASM Whisper is huge, Workers AI uploads audio.

Date: 2026-09-08 08:40 Summary: JS speech-to-text exists in three stacks; Chrome Web Speech is not local, WASM Whisper is huge, Workers AI uploads audio. Visibility: people

[question] js有没有将语音转换为文字的js？

[try to solve]

Context

After dropping the Text Converter Planner ASR cluster, the question is whether JavaScript can transcribe speech at all, and whether that changes the “do not ship STT” product call.

Process

Checked existing direction docs (docs/2026-06-23-10-15-js-media-conversion-feasibility.md, docs/2026-07-28-tool-direction.md A.3, JS package survey). Those list Web Speech API as the STT path and mark audio tools as “usable, not first wave.” They do not split live dictation vs file/YouTube transcription, and they do not warn that Chrome’s SpeechRecognition typically sends audio to Google.

Root cause / analysis

Three different products hide under “JS STT”:

Web Speech API (webkitSpeechRecognition) — live microphone dictation. Chrome/Edge generally use a network recognizer, so a page cannot claim localProcessing the way N1 OCR does. It does not consume an uploaded MP3 or a YouTube URL. Firefox support is historically weak.
On-device Whisper / similar (transformers.js, whisper.cpp WASM) — can transcribe a file in the tab, but English tiny models are still tens of MB; multilingual packs are larger. Same LCP/vendor-git problem as Paddle OCR, worse.
Workers AI @cf/openai/whisper — already typed in this repo’s Worker config. Audio goes to Cloudflare. Same privacy class as optional prompt AI, not same-origin WASM OCR.

Planner heads (speech to text, mp3 to text, youtube to text) match (2) or (3) or a full SaaS, not (1).

Solution

JS can do STT, but not as a drop-in local “mp3/youtube to text” twin of N1. Keep the 2026-09-08 drop for that cluster. A future dictation page would need an honest “Chrome sends audio to the recognizer” label, or a click-to-load Whisper vendor with a POC, or Workers AI + Turnstile — each is a named product decision, not a FAQ on the JPG OCR tool.

Notes / boundaries

Direction A.3 already allowed a simple TTS reader and a recorder as later audio tools. That is not the same as shipping a transcription desk. No page was implemented in this turn.

[actions]

None (docs already listed Web Speech; this log records the local-vs-cloud split)