CrisperWhisper is an advanced variant of OpenAI's Whisper built by the German healthcare-AI company nyra health, and it is designed around a goal most ASR systems quietly ignore: transcribing exactly what was said, word for word. Where standard Whisper cleans up speech into a tidy "intended" transcript — dropping fillers, smoothing over stutters, and normalizing false starts — CrisperWhisper keeps them. The result is verbatim recognition with crisp, precise word-level timestamps, and the project has drawn a steady following on GitHub for filling a gap that clinical, legal, and research transcription workflows genuinely feel. ## What Verbatim Actually Means Here The core distinction is philosophical as much as technical. Intended-style transcription answers "what did the speaker mean to say?" while verbatim transcription answers "what sound did the speaker actually produce?" CrisperWhisper commits to the latter: it transcribes and differentiates fillers like "um" and "uh," preserves repetitions and self-corrections, and marks pauses. For domains where the disfluencies themselves carry signal — speech-pathology assessment, qualitative interview coding, courtroom records — this is not noise to be removed but the data itself. ## Crisp Word-Level Timestamps The headline capability is timestamp accuracy, even in the hard cases around pauses and disfluencies where alignment usually drifts. CrisperWhisper achieves this through an adjusted tokenizer and a custom attention loss applied during training. A later addition beyond the original paper — an AttentionLoss that directly trains the attention scores used for dynamic-time-warping alignment against timestamped data — pushed segmentation performance further. The practical payoff is timestamps that stay glued to word boundaries rather than sliding across gaps of silence. ## Benchmark Standing The model reached first place on the Hugging Face OpenASR Leaderboard for verbatim datasets such as TED and AMI, and overall at the time of publication, and the underlying method was accepted at INTERSPEECH 2024 with an accompanying arXiv paper. Alongside accuracy, the training objective was tuned to mitigate the hallucinated phrases that Whisper-family models sometimes emit on silence or noise — a meaningful reliability gain for unattended batch transcription. ## Getting It Running CrisperWhisper integrates through the paths practitioners already use. It runs with Hugging Face Transformers as a drop-in Whisper-style pipeline, and it also works with faster-whisper for higher-throughput CTranslate2 inference. A bundled Streamlit app provides a click-to-transcribe interface for quick evaluation without writing code, making it straightforward to hear how the verbatim output differs from a conventional transcript on your own audio. ## Trade-offs and Limitations The most important constraint is licensing: CrisperWhisper is released under Creative Commons Attribution-NonCommercial 4.0, so commercial deployment requires a separate arrangement with nyra health rather than free use. Verbatim output is also not always what you want — for clean captions or meeting summaries, the retained "um"s and repetitions add clutter that a downstream cleanup step must remove. And as a Whisper derivative it inherits Whisper's general language coverage and compute profile rather than introducing a fundamentally smaller or faster backbone. ## Who Should Use This CrisperWhisper is a strong fit for researchers and builders who need faithful, precisely time-aligned transcripts: clinicians studying speech disfluency, teams building datasets where filler and pause annotations matter, subtitle pipelines that require tight word timing, and anyone frustrated by Whisper silently rewriting what was actually said. For non-commercial use it is one of the most accurate verbatim options available; commercial users should factor the license into their evaluation.