CrisperWhisper.cpp v1.1.1
This patch release fixes long-form transcription when the same speech appears
more than once near a 30-second chunk boundary.
Fixed
- Timestamp-aware stitching now requires both matching text and matching
absolute time before removing a word at a chunk seam. - Repeated passages are no longer mistaken for duplicate overlap.
- The 55-second five-copy JFK regression now retains all five repetitions:
110 timed words across two chunks, with no invalid ranges, overlaps,
adjacent duplicates, or missing seam passage.
Native ZIPs
- Windows x64 CUDA for RTX 30, 40, and 50 series
- Linux x64 CUDA for RTX 30, 40, and 50 series
- Windows x64 CPU portable
- Windows x64 CPU AVX2
- Linux x64 CPU portable
- Linux x64 CPU AVX2
- Linux ARM64 CPU portable
The Linux packages are built on Ubuntu 22.04. The GNU C++ runtime is linked
statically, and the release rejects binaries requiring GLIBC newer than 2.35.
The ARM64 package has been validated on an Orange Pi 5 Pro running Ubuntu
22.04. The Linux CUDA ZIP bundles the CUDA 12.8 Runtime, cuBLAS, cuBLASLt, and
nvJitLink user-space libraries. It requires a compatible NVIDIA driver but
does not require the CUDA Toolkit, Docker, or virtualization.
Models and license
Model weights are not included in these ZIPs. Download converted GGML models
from drbaph/CrisperWhisper2.0-GGML.
The code in this repository is MIT licensed. CrisperWhisper 2.0 models,
configuration, tokenizers, parameters, and model outputs remain governed by
their own Nyra Health Non-Commercial Research License.
The gated Pro models require their applicable commercial license and access
approval.
Full Changelog: v1.1.0...v1.1.1