✦
Summary
CrisperWhisper 2.0 is Nyra Labs' verbatim speech recognition model for transcribing how people actually speak. It supports controllable output styles, multilingual transcription, word-level timing, and longform audio handling for production speech workflows.
Key Features
Provides controllable verbatim and intended transcription modes.
Supports multilingual speech recognition across many Whisper-supported languages.
Captures fillers, repetitions, stutters, false starts, and vocal sounds.
Produces precise word-level timing for alignment-sensitive workflows.
Handles seamless longform transcription.
Includes Verbatimize workflows for improving legacy speech datasets.
Links to Nyra Labs' public CrisperWhisper repository and model resources.
Targets production-ready inference for real speech data.
About CrisperWhisper 2.0
The system distinguishes between verbatim transcripts that preserve fillers, repetitions, cut-offs, and vocal sounds, and intended transcripts that produce cleaner readable text. It also improves alignment by supervising Whisper-style cross-attention behavior to produce more precise word timings.
CrisperWhisper 2.0 is useful for clinical documentation, conversation analysis, speech datasets, subtitles, podcast workflows, and expressive speech synthesis pipelines. Its controllable style switch helps teams choose whether fidelity to spoken disfluency or readability is more important for a task.