CrisperWhisper 2.0 Large targets verbatim transcription

Nyra Labs has released CrisperWhisper 2.0 Large, an automatic speech recognition model built on OpenAI's Whisper architecture but tuned for a specific goal: transcribing exactly what was said, including the ums, ahs, and false starts that most systems quietly delete.

The model supports English and German, and its headline features are verbatim output with disfluency handling and precise word-level timestamps. That combination makes it a better fit for use cases where fidelity matters more than tidy prose.

Why it matters

Most commercial and open ASR pipelines are optimized to produce clean, readable text, smoothing over hesitations and repetitions. That is helpful for note-taking but a problem for other work:

• Linguistics and speech research, where disfluencies are the object of study
• Legal and medical records that require an accurate account of what was spoken
• Precise alignment tasks like subtitling and audio editing, which lean on word-level timing

By keeping the disfluencies and pinning timestamps to each word, CrisperWhisper 2.0 fills a niche that general-purpose transcription tools tend to ignore. As with any Whisper derivative, teams will want to validate accuracy on their own audio and confirm the licensing terms on the model page before deploying it in production.