Practical overview
When it fits
Use CrisperWhisper when the transcription workflow needs an explicit choice between capturing what was said and producing a cleaner reading version, or when word-level timing is needed.
Who it is for
- Developers building speech transcription workflows.
- Researchers and reviewers who need transcripts that preserve fillers, repetitions, stutters or vocal sounds.
- Teams preparing speech datasets or timed transcripts from audio.
Documented uses
- Transcribe speech while preserving disfluencies and vocal events.
- Create a cleaner, readability-focused transcript of what a speaker meant.
- Attach word-level start and end times to recognized speech or existing text.
- Use audio and a trusted clean transcript to add spoken disfluencies and vocal events.
Requirements & limits
The repository documents installation as a Python package and provides ct2 and transformers extras. Model weights are downloaded from Hugging Face on first load.
- The ct2 installation path is documented for Linux with an NVIDIA GPU; the Transformers path is described for platforms where PyTorch runs, including macOS, Windows and CPU.
- First-time conversion on the ct2 path needs Torch and Transformers, though they can be removed after the converted model is cached.
- The documented model choices include large, turbo, medium and small; turbo is faster with some quality degradation, while medium is described as a size-quality tradeoff.
- The library describes verbatim and intended modes as working across most languages supported by Whisper.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install CrisperWhisper and load a model
Prepare the Python library and its model for transcription.
Before you start
- Python package installation with either the ct2 or transformers extra.
- For the documented ct2 path: Linux and an NVIDIA driver.
- For first-time ct2 model conversion: Torch and Transformers, installed with the convert extra.
Steps
Install
crisperwhisper[ct2]for the documented NVIDIA GPU path, orcrisperwhisper[transformers]where PyTorch runs.For first-time conversion with ct2, install
crisperwhisper[ct2,convert].Load
CrisperWhisperModel()or select a documented size such asturbo,mediumorsmall; the first model load downloads weights from Hugging Face.
Expected outcome
The package and selected model are prepared for the documented transcription calls.
Guide limitations
- The ct2 first-load conversion requires Torch and Transformers; the lean ct2 extra does not install them.
- The Transformers path is described as running wherever Torch does; no specific performance is promised.
Transcribe audio with a selected output style
Create a verbatim or readability-focused transcript, optionally with word timings.
Before you start
- An installed CrisperWhisper package and loaded model.
- An audio file, such as the documented
meeting.wavexample. - A language value for the documented transcription call.
Steps
Load
CrisperWhisperModel()or choose a documented model size.Call
model.transcribe(audio_file, language="en")for the default verbatim output, or setmode="intended"for readable text.Set
word_timestamps=Truewhen word-level start and end times are needed.
Expected outcome
A transcript in the chosen style, with word-level timings when requested.
Guide limitations
- Hotword boosting is documented for Pro models only.
- The README describes multilingual support across most languages supported by Whisper.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- nyrahealth/CrisperWhisper · Official homepage title · retained in its source language