AI for speech as it is actually spoken
nyra labs conducts research, builds speech models and APIs that preserve the details most systems erase — pauses, repetitions, repairs, uncertainty, intent, noise, long-form structure, and change over time.
Understanding the what
Speech recognition that captures every word, exactly as spoken.
CrisperWhisper 2.0 transcribes real, messy human speech — fillers, stutters, self-repairs and pauses — word for word, with state-of-the-art accuracy and timing.
Hover a word for its timestamps — click it to jump there.
CrisperWhisper 2.0
Built for real-world conversations — every stutter, filler, and pause, exactly as spoken.
CrisperWhisper 2.0
Built for real-world conversations — every stutter, filler, and pause, exactly as spoken.
- #1 verbatim transcription accuracy
- Frame-rate timing accuracy
- 25+ languages
Word-level timings
Supervised cross-attention with inference-time sharpening — 36 ms MAE on read speech.
Long-form ready
Conditional continuation removes uncertainty around segment boundaries.
Fastest large Whisper
Speculative decoding inside CTranslate2 — fastest model in its class.
Performance you can verify
Independent evaluations show industry-leading transcription accuracy across languages and challenging real-world speech.
German
DisfluencySpeech · lower is betterEnglish
DisfluencySpeech · lower is betterLatest releases
CrisperWhisper 2.0
World's #1 verbatim speech recognition — every filler, pause and repair, exactly as spoken.
The Verbatim Speech Benchmark
Sixteen ASR systems scored on the fillers, repetitions and cut-offs that make speech real — 20+ typed metrics, not one WER.
Word-level timing from attention
Word boundaries to within 30 ms — supervised straight from the cross-attention heads that already lean toward alignment. No separate aligner.