Controllable speech recognition for verbatim and intended transcripts with word timings.
Switch between exact disfluencies and clean intended text with precise word-level timestamps.
CrisperWhisper 2.0 is an open-source speech recognition tool that lets you choose between strict verbatim transcripts—including every filler, stutter, and vocal sound—and clean intended text. It is built for researchers, linguists, and developers who need precise control over audio transcription and word-level timings.
The model provides two distinct transcription modes from a single recording: verbatim (capturing every um, pause, and repetition) and intended (optimized for readability with formatted dates and numbers). It also includes a verbatimize function to upgrade clean transcripts with actual audio disfluencies and supports CTranslate2 production inference with speculative decoding.
It suits researchers, speech pathologists, and developers building transcription pipelines who require granular control over disfluency tracking. It does not suit casual users looking for a plug-and-play desktop meeting recorder without command-line interaction.
Switch between exact verbatim output with all disfluencies or clean intended text from a single audio file.
Achieve high-precision word-level boundary errors using supervised cross-attention or an optional forced aligner.
Inject actual audio disfluencies and vocal events into an existing clean text transcript.
Run on a CTranslate2 runtime featuring speculative decoding and built-in looping-hallucination mitigation.
The software code is free under its repository terms, with standard models subject to a non-commercial research license. Pro models and commercial deployment require getting in touch with the creators for a commercial license.
873 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Verbatim mode writes down exact speech including fillers, stutters, and vocal sounds, while intended mode outputs a clean, readable version with properly formatted numbers and dates.
Standard models use a non-commercial research license and require contacting Nyra Labs for commercial licensing, while Pro models are strictly commercial license only.
Given an audio file and a trusted clean transcript, verbatimize reproduces the content word-for-word while inserting the exact disfluencies and vocal events present in the recording.
See Zekai first in Google