On January 6, the release of 1.7.4 high-performance automatic speech recognition system took place. whisper.cpp, implementing the model Whisper from OpenAI, written in C and C++ and distributed under the MIT license.
The project utilizes a custom machine learning tensor library GGML, written in C and optimized for various backends (CUDA, Vulkan, BLAS, SYCL, OpenCL, etc.).
Change log:
- Renaming of executable files:
main -> whisper-cli bench -> whisper-bench stream -> whisper-stream command -> whisper-command server -> whisper-server talk-llama -> whisper-talk-llama
- Examples talk and talk.wasm have been removed.
- An option to suppress non-speech tokens has been added to whisper-server.
- The suppress_non_speech_tokens parameter has been renamed to suppress_nst.
- The API and utilities now include an option no_speech_thold to set the probability of speech absence at the segment level during transcription.
- The option —suppress_nst has been added to whisper-cli.
- Other improvements and bug fixes.
Source: linux.org.ru
