Repository navigation
Cortiq 0.8.3
Whisper large-v3 and large-v3-turbo
- Convert supported 128-bin Hugging Face Whisper checkpoints to CMF with
cortiq convert. - Transcribe WAV PCM/float audio with
cortiq transcribeon CPU, Apple Metal, or wgpu/Vulkan. - The mixed
q4tpprofile retains attention and the shared decoder token/output matrix inq8_2f, with feed-forward weights inq4tp.
Whisper weights are not included. Validate word error rate on your target audio before production use. See the README for conversion and transcription examples.
Validation
- Workspace build and tests passed on Linux and macOS.
- Clippy, CLI MSRV (Rust 1.88), and Vulkan-feature compile checks passed.
- Release CLI built on the RTX PRO 4000 Linux pod;
cortiq verifyand English/Russian transcription smoke tests passed.