CrisperWhisper 2.0 brings controllable verbatim speech recognition

CrisperWhisper 2.0 is an open-source speech-to-text model family from nyra labs that fixes one of the most annoying quirks in automated transcription: deciding whether you want an exact word-for-word record or clean, readable prose.

Most existing models randomly blur the line, keeping a filler word in one sentence and dropping it in the next. CrisperWhisper 2.0 turns that behavior into an explicit switch.

What it does

Flip on Verbatim Mode, and the model writes down exactly what happened in the audio. That includes every stutter, repetition, false start, hesitation, and vocal event like laughter. It is built specifically for clinical analysis, legal work, or training speech models where disfluencies carry real information.

Switch to Intended Mode, and it strips away the ums and uhs, cleans up false starts, and formats numbers, dates, and email addresses into polished, publish-ready text.

Its word-level timing is also remarkably sharp. Standard Whisper alignment can drift by hundreds of milliseconds, but CrisperWhisper 2.0 logs a mean boundary error of just 29.6 ms on read speech and 40.6 ms on spontaneous conversation. That puts it ahead of commercial alternatives like ElevenLabs Scribe v2 and xAI Grok STT on benchmarked timing accuracy.

To handle long audio without the classic Whisper looping bug, it ditches brittle timestamp tokens in favor of conditional continuation. By passing the tail end of the previous transcript into the next processing window, it preserves continuity and avoids hallucinations at the seams.

There is also a tool called Verbatimize. Feed it an audio track alongside any existing clean transcript, and it will retroactively restore all spoken hesitations and laughter, making it vastly cheaper to build expressive TTS datasets.

How to run it

nyra labs released four model sizes supporting multilingual transcription across most languages supported by Whisper.

The smallest version sits at 0.2B parameters and weighs less than 500MB. You can run it locally on almost any basic consumer laptop without needing a dedicated GPU.

The largest flagship model has 2B parameters and takes up roughly 3GB of storage, which fits easily inside standard consumer GPUs. It ships with a custom CTranslate2 runtime optimized for speculative decoding and low-latency production serving.

You can test it immediately on their free Hugging Face Space by uploading an audio clip, or pull the weights and instructions from Hugging Face for local setups.

Why it matters

If you work on subtitle generation, podcast editing, or speech dataset curation, this is an easy recommendation. It cleanly separates verbatim accuracy from formatted readability instead of forcing a messy compromise.

nyra labs took a very pragmatic route here. Rather than bloating model scale, they focused on boundary precision, hallucination suppression, and operational efficiency. Delivering 30 ms word alignment in a sub-500MB package gives developers a genuinely competitive open-source alternative to pricey proprietary transcription APIs.