whisper.cpp v1.9.4: Windows on ARM support and ggml update to 0.23.0

Version v1.9.4 of the whisper.cpp launcher has been released. In the GitHub repository, the changes have been published: Windows on ARM support has been added to the build pipeline, the ggml backend has been updated to version 0.23.0, repeated decoder startup has been fixed, and the call order of the language auto-detection callback has been corrected.

What it means

For those who run whisper.cpp on their own machine, the release brings several practical changes. The Windows on ARM build is now part of the release process — this concerns devices with ARM processors running Windows. The ggml update to 0.23.0 affects the compute backends: in Metal, missing tunings for M3 Max have been added; in SYCL, the rms_norm+mul+add and add+add operation chains have been merged; and an uninitialized variable in q5_1 quantization for s390x has been fixed. The change in the call order of encoder_begin_callback before language auto-detection may affect server mode: the response to a detection request now returns the language. Reinitializing the decoder between calls eliminates instability during consecutive runs. The developer does not specify memory requirements or particular models for this release. Link to the news feed about AI tools for 3D and video: cpu3d.com/ainews.