Confucius4-R2T2 in your browser

2B speech recognition (Chinese, English and 28 more languages) running on WebGPU. Audio never leaves the device.

idleModel not loaded (~1 GB, cached after the first visit)

Upload audio

examples

Transcript appears here. Load the model, then pick an example, upload a file or start the live mic.

Settings

Quantization Language Context / hot words (optional system prompt) Changing quantization reloads the model. Live mic uses the R2T2 rolling 16 s window: text in ink is final, grey text may still change.