source separation

Source separation isolates individual audio sources from mixed recordings, like extracting vocals from a song. **Use cases**: Extract vocals (karaoke creation), isolate instruments, remix production, audio restoration, podcast cleanup, music transcription. **Approaches**: **Spectrogram masking**: Predict time-frequency masks for each source, apply to spectrogram, invert. **Waveform-based**: End-to-end models directly output separated waveforms. **Hybrid**: Operate on both domains. **Key models**: Demucs (Meta, state-of-art), Spleeter (Deezer, fast/simple), Open-Unmix, BSRNN. **Common separation tasks**: Vocals/accompaniment (2 stems), vocals/drums/bass/other (4 stems), full instrument separation. **Technical details**: U-Net architectures, multi-scale processing, trained on synthetic mixtures with known components. **Quality metrics**: SDR (Signal-to-Distortion Ratio), SIR, SAR. **Challenges**: Overlapping frequencies, artifacts in separated sources, generalization to diverse music. **Tools**: Demucs CLI/Python, Ultimate Vocal Remover (GUI), online services. **Applications**: Sampling/remixing, cover versions, music education, accessibility. Powerful creative tool for audio professionals.