v1.1.1

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

v1.1.0

CrisperWhisper.cpp v1.1.0

Native C++/ggml inference for the public CrisperWhisper 2.0 models.

Highlights

• Optional supervised cross-attention word timestamps
• Overlap-safe long-form stitching with duplicate and invalid-timestamp protection
• Windows x64 CUDA package for RTX 30, 40, and 50 series GPUs
• Windows and Linux x64 CPU packages
• Experimental Linux ARM64 portable CPU package
• Automatic WAV, MP3, FLAC, and Ogg decoding, mono downmix, and 16 kHz resampling
• Bundled JFK validation sample

Linux ARM64 compatibility

The ARM64 asset was rebuilt on native Ubuntu 22.04 ARM hardware after an Orange Pi test exposed an Ubuntu 24.04 ABI dependency in the original upload.

The corrected crisperwhisper-v1.1.0-linux-arm64-cpu-portable.zip:

• Requires no newer than GLIBC 2.34, so it is compatible with Ubuntu 22.04 GLIBC 2.35 and newer releases such as Ubuntu 24.04
• Statically links the GNU C++ runtime and has no GLIBCXX runtime-version dependency
• Passed native ARM64 unit tests
• Loaded the Turbo GGML model and completed a timestamped JFK transcription on the native ARM64 runner
• Does not contain model weights

No Docker or virtualization is required to use any release archive.

Models

Download GGML model files separately from https://huggingface.co/drbaph/CrisperWhisper2.0-GGML.

The CrisperWhisper 2.0 model files remain governed by Nyra Health's Non-Commercial Research License: https://huggingface.co/nyralabs/CrisperWhisper2.0_large/blob/main/LICENSE.md.

Full Changelog: v1.0.0...v1.1.0