Whisper Timestamp Studio

1 Audio

Choose a file

No audio loaded yet — record something, upload a file, or load the example.

2 Settings

Model Downloaded once, then cached by your browser. Language Auto-detect reads it off the first 30 s of audio. Task Translate always outputs English. Timestamp granularity Word timing is slower; the app falls back to segments if unsupported.

3 Transcribe

Waiting for audio…

4 Results

Transcript

Timestamped segments

Nothing transcribed yet.

Subtitles

The SRT preview appears here.

Run details

Nothing yet.

How it works

The page loads a Whisper checkpoint exported to ONNX from the Hugging Face Hub and runs it with Transformers.js. Audio is decoded and resampled to 16 kHz mono locally, decoded in Whisper's native 30-second windows, and the timestamps stitched back together. On WebGPU the encoder runs in fp32 and the decoder in 4-bit; without WebGPU it falls back to WASM with int8 weights.

Timestamps come from Whisper's own decoding. Word-level timing needs a *_timestamped export — every model in the dropdown is one of those.