nyrahealth/CrisperWhisper

Built on OpenAI's Whisper, CrisperWhisper employs a custom tokenizer and attention loss mechanism during training to achieve precise word-level timestamp alignment, particularly around disfluencies and pauses. It integrates seamlessly with both 🤗 Transformers and Faster Whisper pipelines, enabling deployment in existing speech recognition workflows. The model prioritizes verbatim transcription including fillers ("um", "uh") and speech artifacts, ranking first on the OpenASR Leaderboard for verbatim datasets like TED-LIUM and AMI.

927 stars. No commits in the last 6 months.