8000 Release 1.3.0 - GPU Support · Macoron/whisper.unity · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

1.3.0 - GPU Support

Compare
Choose a tag to compare
@Macoron Macoron released this 30 Nov 21:54
· 15 commits to master since this release
25c8d26

This release introduce whisper.cpp update to 1.5.1, GPU inference support and other minor improvements.

Whisper.cpp updated to 1.5.1

whisper.cpp 1.5.1 got a lot of improvements and bug fixes including better GPU usage.

Check original release notes for more information.

GPU Support

Whisper now supports GPU acceleration. This can drastically improve performance for some hardware.

Model CPU CUDA
tiny 1188 ms 185 ms
small 8992 ms 517 ms
large-v2 60325 ms 1946 ms

Tests of "jfk.wav" transcribing on Windows with Intel Core i5-12400F and Nvidia Geforce RTX 2070 Super.

Model CPU Metal
tiny 1113 ms 189 ms
small 6319 ms 860 ms
large-v2 40608 ms 3888 ms

Tests of "jfk.wav" transcribing on Apple M1 Pro.

For Windows and Linux you would need Nvidia GPU and installed CUDA Toolkit (tested with 12.2.0). Unity project compiled with enabled CUDA expects your end-users to have Nvidia GPU and CUDA libraries. Trying to run build without it will result error.

For MacOS you would need ARM CPU, like M1 or newer. iOS Metal inference isn't yet supported. In case of Intel or older hardware, whisper.cpp should fallback to CPU inference.

To activate GPU inference, go to Project Settings => Whisper => Enable CUDA or Enable Metal. For more information, check README.

Other

Ubuntu libs now compiled on Ubuntu 20.04. This might cause problems with Ubuntu 18.04. If you need support for earlier versions of Ubuntu or other distros, consider recompiling libs from source.

New loop mode for microphone was added. It creates a new endless non-stopping stream using Unity build-in circular microphone loop. This is very useful for whisper streaming transcription. To activate it - set Loop in MicrophoneRecord to "true".

What's Changed

Full Changelog: 1.2.1...1.3.0

0