Models
Open models for real human speech
Verbatim recognition today; timing, diarization, analysis, and speech tokens as the research program lands them. Weights on Hugging Face, benchmarks you can rerun.
consultation_recording.wav0:42 · de-noised · verbatim
CrisperWhisper 2.0I'd[uh]liketobookanappointmentformy[lipsmack]mybackpain.
appointment · 1.42 – 2.05 s· word-level timestamps included
Run it
from transformers import pipeline
asr = pipeline(
"automatic-speech-recognition",
model="nyrahealth/CrisperWhisper",
)
out = asr("consultation.wav",
return_timestamps="word")Open weightshuggingface.co/nyrahealth/CrisperWhisper
Free for research. For products, licensing takes minutes — model license.
Weights on Hugging Face
Our released models ship as open weights with model cards — download them, fine-tune them, ship them.
Benchmarks you can rerun
We publish the benchmarks we measure ourselves against, with the data to reproduce every number.
Licensing for products
Research use is free. Putting a model in a commercial product takes one short license — not a sales cycle.
All models
- CrisperWhisper 2.0ShippedVerbatim speech recognition that keeps fillers, repairs, and pauses — with word-level timestamps.
- nyra forced alignerComing soonPhoneme-level timing, robust to spontaneous, disfluent, real-world speech.
- PINTPublishedOur parallel-invariant tokenizer: the content of speech, with everything else factored out.
- Fully factorized codecsIn progressContent, speaker, prosody, and conditions — each in its own place, at its own granularity.
Put verbatim speech in your product
Run CrisperWhisper 2.0 in your browser, pull the weights from Hugging Face, or talk to us about a commercial license.