3 Key Ways to Improve Demucs Separation Accuracy
The quality of the audio source and the choice of parameters directly affects the separation, and the following are best practices:
1. Selection of an optimal model
- Priority use
-n htdemucs_ft
Fine-tuning the model (version v4) - Complex music to try
-n htdemucs_6s
Six-track modeling to separate more instruments - Classical music is recommended
-n mdx_extra
Specialized AI
2. Optimizing input quality
- Use lossless WAV format instead of MP3 to reduce compression loss
- Ensure audio sampling rate ≥ 44.1kHz
- Remove the muted portion of the audio front end
3. Parameter fine-tuning techniques
- increase
--float32
Improve precision with 32-bit floating point calculations - Avoid using
--segment
Parameters cut too small segments (≥10 seconds recommended) - pass (a bill or inspection etc)
--shifts=2
Increase the number of predictions (which will increase processing time)
Tests have shown that using the WAV+htdemucs_ft combination can improve the separation accuracy by about 151 TP3T over the default setting.
This answer comes from the articleDemucs: free open source tool for separating music tracksThe