FACEBUBBLE / LOCAL EXPERIMENT

On-device captions lab

Compare English speech recognition in your browser. No room connection, audio upload, recording storage, or speaker playback.

Recognition engine

Both use single-thread WASM CPU in a worker. Whisper is not a streaming model. WebGPU is not enabled in this baseline. Models are downloaded from this host when requested, not on page load. Allow roughly 200 MB for Sherpa and 70 MB for Whisper plus runtime. Mobile speed and memory are unverified.

Pace sample/file like live audio (real-time latency)

Or choose an audio file (up to 10 MiB, 30 seconds)

Microphone starts only after your click and permission. Maximum 30 seconds. Stop releases capture immediately and cancels pending inference. No camera is requested. Transcripts are ephemeral and can be wrong.

Choose a model, sample, or file. Microphone is off.

Transcript

Last measurement

Unpaced mode measures throughput, not conversational latency. Paced mode delivers 200 ms of audio at a time on its real-time schedule. First text excludes model loading and audio decode and includes each engine's chunking policy. Loading is reported separately. Sample is the public JFK speech fixture used by Transformers.js. No audio is played.

No measurement yet.