nyrahealth/CrisperWhisper

What it solves

CrisperWhisper 2.0 provides high-accuracy, verbatim speech-to-text transcription. It solves the inconsistency in standard ASR systems that struggle to decide whether to transcribe exactly what was spoken (including stutters and fillers) or the intended meaning (clean text). It also addresses common ASR issues like word-timing inaccuracies, chunk-boundary artifacts in long audio, and looping hallucinations.

How it works

The system uses a CTranslate2 runtime with speculative decoding for speed and a specialized decoder to mitigate hallucinations. It employs supervised cross-attention alignment to achieve precise word-level timings. For long-form audio, it uses "conditional continuation" to ensure seamless transitions between transcription windows. It also includes a "Verbatimize" feature that can take a clean transcript and an audio file to insert the actual disfluencies present in the recording.

Who it’s for

Developers and researchers needing precise, word-for-word transcription for clinical speech analysis, TTS dataset construction, or production-grade ASR where controllable style (verbatim vs. intended) is required.

Highlights

• Dual-Mode Transcription: Explicitly switch between "verbatim" (including fillers and stutters) and "intended" (clean, formatted text).
• High Precision Timings: Word-boundary error as low as 29.6 ms on read speech.
• Verbatimize: Upgrades clean transcripts into verbatim ones by inserting real vocal events from audio.
• Seamless Longform: Transcribes audio of any length without duplicated or dropped words at window seams.
• Production Ready: Includes speculative decoding for 1.3-1.4x speedup and built-in hallucination mitigation.