Speech recognition that stays off the server

What it does

Vosk-browser wraps the Vosk speech recognition engine in a WebAssembly build purpose-built for browser WebWorkers. You load a model (a .tar.gz file), wire it to microphone input through the Web Audio API, and get text back — all without a network round-trip to a speech API.

The interesting bit

The heavy lifting isn’t new; Denis Treskunov did the original WebAssembly port of Kaldi. This project packages an updated Vosk build and, crucially, makes the browser integration ergonomic. The WebWorker constraint matters: it keeps the main thread free while the recognizer chews on audio, which is the only sane way to do real-time ASR in a browser without jank.

Key highlights

• 13 languages supported in the live demo
• CDN loadable via jsdelivr as a global Vosk object
• API surface is small: create model, instantiate KaldiRecognizer, feed it AudioBuffer chunks via acceptWaveform()
• Explicitly browser-only; NodeJS users are directed to official bindings
• Examples and API docs live in ./lib/README.md and ./examples/

Caveats

• No tests yet (listed in Todos)
• Model files are your problem to host and serve; no cloud model delivery
• Speaker identification models are not yet demonstrated

Verdict

Worth a look if you need offline, privacy-preserving speech-to-text in a web app and can stomach shipping multi-megabyte model files. If you want turnkey cloud accuracy or NodeJS server-side ASR, this isn’t your tool.