Speech to Text

Speech recognition built from the ground up for streaming. Connect a WebSocket, send audio, and transcript segments stream back word by word with per-word timestamps.

3 supported modes to support your use case:

• Instant - Stream audio in and words are transcribed as soon as confidence allows. Sub 200ms word to transcript latency in our own testing.
• Demand - Stream audio in and signal when a transcript is requested, use your own systems to detect turn boundaries, critical conversational moments, or at regular intervals.
• Transcribe - Traditional “offline” transcription mode. Upload audio in up to 5 minute segments all at once, get the whole transcript back.

Currently supporting English only for transcription.

We are also launching with this model some built in capabilities for you to try out:

• Keyword Boosting - submit with your api request a list of up to 100 keywords and we will use them to bias the models output transcriptions to help with domain specific words.
• PII Redaction - US centric PII redaction supported for Demand and Transcribe mode. We currently look for payment card numbers, Social Security numbers, phone numbers, email addresses, dates of birth, account and ticket numbers, and ZIP codes. Detected spans are filtered from the output transcriptions and can optionally be handed back separately in the response for special handling.

This model and additional features are experimental and in a Beta state, so expect some bugs and issues, and expect new features to appear regularly. Please feel free to try everything out as much as you want while we work to improve it. If you have specific feedback for the model development team please use the feedback form to submit it.