Wav2Vec2-BERT STT

Speech-to-text inference · 16 kHz mono · peak-normalized to [-1, 1].

Server status: checking…

  

REST upload (audio ≤ 15s)


  

WebSocket live streaming

disconnected

VAD segments speech into utterances. Partials update every ~1 s while you speak; a final is emitted when you pause (or on Stop & finalize). Junk finals (a single stray letter) are dropped after decoding, so short real words like «за» come through. Audio is resampled to 16 kHz mono and peak-normalized to [-1, 1] before transcription. Toggle Denoise any time — it applies to the next utterance, and it lowers accuracy on this model. Toggle KenLM any time too: on = beam search with the language model (default, better accuracy), off = plain greedy decode.

Partial:

    
Final: