Kyutai STT
Kyutai
Kyutai STT is a streaming speech-to-text model architecture from the Paris open-science lab Kyutai, released as two checkpoints: stt-1b-en_fr (~1B parameters, English and French, a 0.5 second delay, and a built-in semantic voice activity detector) and stt-2.6b-en (~2.6B parameters, English only, a 2.5 second delay). It is a decoder-only Transformer that consumes Mimi audio tokens at 12.5 Hz and emits a text stream time-aligned to the audio, producing punctuated transcripts with word-level timestamps. Weights are released under CC-BY 4.0.
Strengths
- Streaming transcription with a 500ms text delay on kyutai/stt-1b-en_fr, so words appear 500ms after they are spoken
- Built-in semantic voice activity detector predicts the probability that the speaker is done talking, with a delay that adapts to content and intonation
- Outputs well-formatted transcripts with capitalization and punctuation plus word-level timestamps
- High batching throughput: 400 real-time audio streams simultaneously on a single H100
- Production Rust websocket server serves 64 simultaneous connections at a 3x real-time factor on an L40S
- Three first-party implementations (PyTorch, Rust, MLX) including on-device inference on iPhone and Mac
- Reported to perform well on audio as long as 2 hours with no additional changes
Caveats
- Language coverage is tiny: stt-1b-en_fr handles only English and French, and stt-2.6b-en is English-only, against the dozens covered by Whisper-class models.
- Word error rate is published only as a chart on the project page; no numeric benchmark score is stated as text anywhere on Kyutai's own pages, so the accuracy claims cannot be verified from the vendor.
- The semantic VAD, the model's headline feature, is available in the Rust server but not yet in the other implementations.
- The more accurate 2.6B checkpoint pays for its accuracy with a 2.5 second delay, five times the 1B model's 0.5 second delay.
- No vendor-hosted API and no stated release date; users self-host, and the CC-BY 4.0 weights licence carries an attribution obligation.
- Vendor
- Kyutai
- License
- CC-BY 4.0
- Modalities
- audiotext
Capabilities
Resources
Pricing
No vendor pricing is published: the weights are released under CC-BY 4.0 and Kyutai provides PyTorch (research), Rust (production) and MLX (Apple silicon) implementations for self-hosting via the delayed-streams-modeling repository.