On June 17 and 19, versions 1.9.0 and 1.9.1 of the high-performance automatic speech recognition system were released. whisper.cpp, implementing the model Whisper from OpenAI, and a tensor-based machine learning library GGML and binary format GGUF.
The libwhisper library, demos, and console utilities are provided: whisper-bench, whisper-cli, whisper-command, whisper-lsp, whisper-quantize, whisper-server, whisper-stream, whisper-vad-speech-segments, parakeet-cli, and parakeet-quantize.
The projects are written in C and C++ and are distributed under the MIT license.
Changes:
- The parakeet-cli console utility has been added, supporting the model NVIDIA Parakeet;
- NVIDIA Parakeet support has also been added to the Ruby wrapper.
Some important changes made in intermediate versions since the news whisper.cpp 1.8.0:
- C++ exception handling in whisper_init_with_params_no_state function;
- ffmpeg support has been completely rewritten (without using GPL code taken from ffmpeg examples), and the usage of ffmpeg in whisper.cpp has been clarified;
- The whisper-cli utility now has a new option —version to display the version.
Source: linux.org.ru
