CrisperWhisper

CrisperWhisper is an advanced fine-tune of Whisper Large by Nyra Health aimed at true verbatim transcription — every stutter, false start, filler and pause is captured rather than smoothed over as standard Whisper does.

• Encoder–decoder Transformer (Whisper Large v2/v3 base, ~1.5B params)
• Adjusted tokenizer with explicit filler/pause tokens
• Custom attention loss during fine-tuning to align attention to word boundaries, enabling crisp timestamps even around disfluencies

• Verbatim output: transcribes [UM], [UH], stutters, repetitions, and false starts
• Word-level timestamps with substantially higher F1/IOU than Whisper Large v2/v3 on AMI, Common Voice, and TIMIT
• Hallucination mitigation for cleaner output on noisy or short audio
• Available in both Transformers and Faster-Whisper (CTranslate2) formats

Medical scribing, linguistic/phonetic research, discourse analysis, legal transcription, dubbing/subtitling, and any pipeline that requires exactly what was said.