Fish Audio transcribe-stream-1 is a preview realtime speech-to-text model built for voice agents. It performs turn detection, backchannel recognition, background awareness and history-aware transcription over a WebSocket, ranked #1 of 20 on Cekura’s STT benchmark at 2.21% word error rate. The preview is available now via the Fish Audio API.
Fish Audio’s transcribe-stream-1 is a realtime speech-recognition model accessed through a WebSocket. It streams draft text and commits a final transcript for each turn, with turn signals intended to help voice agents distinguish a user’s turn from acknowledgements, background sound, or interruptions. The service also supports conversation history and custom vocabulary as transcription context. Fish Audio documents use cases including voice agents, live captions, voice commands, and phone calls. Its benchmark page reports results across public clips and licensed phone-call recordings.
Be the first Cosign
Add it to your Products I use list. You can also share why you use it.
