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.