What's Changed
- Server: new WebUI with glassmorphism design (light/dark theme), microphone recording, audio trimmer, login authentication
- Audio: in-memory audio decoder with FFmpeg and miniaudio backends
- CLI/Server: add --backend/--cpu/--cuda backend selection options
- KV Cache: support asymmetric quantization with separate K/V cache types
- Quantize: add CosyVoice3-2512 quantization profiles, tensor map support
- CLI: Ctrl+C to stop/restart playback in interactive mode, second press to force quit
- Frontend: fix voice cloning similarity issue
- Audio: fix FFmpeg backend resampling delay residual issue
Full Changelog: b8b69f7...0aaa9ef
Release Packages
These packages only include the core cosyvoice binaries and tools, plus ONNX Runtime.
GGML, ICU, and (for FFmpeg backend) FFmpeg are not included.
To use cosyvoice, you need to download the matching backend libraries separately:
-
GGML:
Download fromllama.cppreleases and extract the GGML library.
Place GGML files in the same directory ascosyvoicebinaries. -
ICU:
Required unless using a "No ICU" build.
On Linux/macOS: install via your package manager.
On Windows: Download ICU. -
FFmpeg:
Required only for FFmpeg backend builds (not needed for miniaudio).
On Linux/macOS: install via your package manager.
On Windows: Download FFmpeg
For Windows, use stable n8.1 builds, e.g.
ffmpeg-n8.1-latest-win64-lgpl-shared-8.1.zip.
Windows
Windows x64 (FFmpeg)
Windows x64 (miniaudio)
Windows x64 (miniaudio, No ICU)
Windows x64 (No Frontend, No Audio, No ICU)
Linux
Ubuntu x86_64 (FFmpeg)
Ubuntu x86_64 (miniaudio)
Ubuntu x86_64 (miniaudio, No ICU)
Ubuntu x86_64 (No Frontend, No Audio, No ICU)
macOS
macOS arm64 (FFmpeg)
macOS arm64 (miniaudio)
macOS arm64 (miniaudio, No ICU)
macOS arm64 (No Frontend, No Audio, No ICU)
Notes
- Linux/macOS: Install ICU and FFmpeg (if needed) from your system package manager.
- Windows: Download ICU from unicode-org/icu releases.
- FFmpeg can be downloaded from: https://github.com/BtbN/FFmpeg-Builds/releases
- For FFmpeg on Windows, use the stable n8.1 builds, for example:
https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-n8.1-latest-win64-lgpl-shared-8.1.zip - Known issue with pre-built GGML CUDA backend: Some users have reported noise in generated audio when using pre-built GGML binaries from llama.cpp releases with the CUDA backend. I testing confirmed this issue with pre-compiled GGML CUDA builds, while self-compiled GGML from source did not exhibit the problem. If you encounter noise when using the CUDA backend with pre-built GGML, we recommend building both this project and GGML from source as a workaround. Refer to the Build section for instructions.