Source Separation and Noise Cancellation

  • Cocktail party problem: separating overlapping sources
  • Classical methods: ICA, NMF, beamforming (delay-and-sum, MVDR)
  • Deep learning methods: deep clustering, Conv-TasNet, DPRNN, SepFormer
  • Permutation invariant training (PIT)
  • Music source separation: Demucs, Open-Unmix
  • Active noise cancellation (ANC): feedforward vs feedback, adaptive filtering (LMS, NLMS)
  • Noise reduction and speech enhancement: spectral subtraction, Wiener filtering, neural speech enhancement
  • Echo cancellation: acoustic echo cancellation (AEC), double-talk detection