This release introduce whisper.cpp update to 1.7.5, Vulkan inference support and other minor improvements.
Whisper.cpp updated to 1.7.5
In recent versions, whisper.cpp has undergone significant changes to its compiled library structure, making it more modular. While the Unity bindings logic remains largely the same, most of the code related to compilation and native library handling has been completely rewritten.
Version 1.7.5 also improves performance on some hardware. In future releases, it makes possible to support loading hardware-compatible libraries (e.g., non-AVX), allowing whisper.cpp to choose the optimal configuration at runtime.
Better GPU Support
Windows and Linux now support Vulkan for GPU acceleration. Vulkan does not require any additional software to function, making it ideal for more robust software distribution.
| Model | CPU | Vulkan |
|---|---|---|
| tiny | 1118 ms | 90 ms |
| small | 8499 ms | 305 ms |
| medium | 27469 ms | 721 ms |
Tests of "jfk.wav" transcribing on Windows with Intel Core i5-12400F and Nvidia Geforce RTX 2070 Super.
iOS now supports Metal. Note that whisper.cpp supports Metal only on Apple7 GPUs or newer (starting from iPhone 12). On older hardware, inference will fall back to the CPU.
To activate Metal or Vulkan, simply enable the Use GPU option in the WhisperManager settings.
CUDA is no longer supported and was replaced by Vulkan. The compiled CUDA library exceeds 300 MB in size, which no longer fits in the Git repository. It may return in future releases. If you require CUDA support, please use an earlier whisper.unity releases.
What's Changed
- Fixed loading model from path with non-English letters by @Macoron in #98
- Whisper 1.7.5 + Vulkan support by @Macoron in #107
Full Changelog: 1.3.2...1.4.0
