Nothing is noise
Speech carries more than words. nyra labs conducts research and builds open models that preserve what was said and measure when, who, and how - so real human speech can be understood in full.
CrisperWhisper 2.0: Every word. Every pause. Precisely timed.
CrisperWhisper 2.0 is a speech-to-text model that transcribes speech as it actually happened - with fillers, repetitions, false starts, and vocal sounds included. It is the first released layer of our research program.
Real CrisperWhisper 2.0 output. Select any word to play from there.
Nothing gets lost
Word-level timestamps and best-in-class verbatim transcription accuracy across languages.
- #1 verbatim transcription accuracy
- #1 word-timing accuracy
- 10+ languages
Word-level timing
The most accurate word-level timestamps on the market: 30 ms word boundaries via guided cross-attention.
Recover the verbatim layer
Turn existing spontaneous-speech corpora into faithful verbatim data with the Verbatimize task, while preserving the transcripts you already trust.
Open, fast, and production-ready
Open weights and code on Hugging Face, seamless longform, stronger hallucination resistance, and optimized inference make the model ready for real applications.
Performance you can verify
Open evaluation shows industry-leading transcription accuracy across languages and challenging real-world speech.
View the open benchmarkGerman
Disfluency F1 · higher is betterEnglish
Disfluency F1 · higher is betternyra forced aligner: Every word and every laugh, exactly in time.
Give it audio and a verbatim transcript, and it returns when every word, filler, laugh and cut-off happened. It is made for CrisperWhisper transcripts and stays accurate where other aligners break: next to laughs and in noise.
20.6 ms
Mean word-boundary error on conversational speech, 19% below the best of ten baselines, with 2.7× fewer misalignments over 200 ms than MFA.
Every sound gets its place
Fillers, laughs and cut-offs are units of the model, so they get timestamps of their own instead of being smeared into the words around them.
Real output of the published model on a clip from the DisfluencySpeech corpus, which is part of its training data. Select a word to zoom in on its sounds, then pour in noise.
Use them your way
Run the weights yourself, call the API, or license a model for your product.
Run it yourself
Weights and code for every model, free for research and other non-commercial use.
Call the API
CrisperWhisper 2.0 at $0.90 per audio hour. Top up from $5, with no subscription.
Our research program
From spoken detail to speech understanding
Our program follows a deliberate order: preserve speech, measure its structure, separate the factors it carries, and build models that can learn from the whole signal.
How precisely do speech systems time words?
How precisely do speech systems place word boundaries? An open benchmark that scores 22 forced aligners and ASR systems against 100k hand-labelled words on conversational, stuttered and noisy speech, and measures how fast each one runs.
The nyra forced aligner
Word- and phoneme-level forced alignment for real speech: noise-robust WavLM features, explicit units for fillers and vocal sounds, and verbatim training data. 20.6 ms on conversational speech, 2.7× fewer catastrophic errors than MFA.
Work with the lab
Research collaborations, data partnerships and shared evaluation. Tell us what you are working on.
Get in touch