bandit-v2-multi-mlx
audio track separation for MLX
1 min read
The model converts the official bandit-v2 multilingual weights to work with the MLX framework (Swift / Python) on Apple Silicon chips. It performs cinematic audio separation into individual tracks: speech, music, and sound effects. The conversion was performed without quantization or retraining, and the tensors are formatted in safetensors. The tool is useful for video editors, sound engineers, and macOS audio application developers.
Key features
- Support for multilingual audio separation (speech, music, sound effects)
- Safetensors format for native operation via MLX
- Without quantization and loss of original weight quality
- Processing of mono audio at 48 kHz in 8-second chunks
What you can do
- Separate a movie or video audio track into clean speech, background music, and sound effects
- Integrate the model into local macOS audio applications based on the MLX framework
- Clean dialogues from extraneous noise and music for post-production
Example
The model accepts mono audio at 48 kHz as input (channels are processed separately) and processes it in 8-second fragments with overlap for seamless stitching of results.
Read more
DetailsThe model is based on the work of bandit-v2 (Netflix, arXiv:2407.07275).
Model size: about 40 million parameters (40MB).
Sample rate: 48 kHz, n_fft 2048, hop 512.
License: CC BY-SA 4.0.










Comments