Solving Audio Inverse Problems with Time Domain Losses

This project was on solving audio inverse problems where time domain signal is lost or corrupted (ex. audio declipping or declicking). We have proposed a maximum likelihood estimation algorithm that relies on nonnegative tensor factorization (NTF) for probabilistic modeling of the time frequency domain audio samples.

 

We soon realized that this algorithm can in fact be used for many other problems, such as source separation with audio inpainting, informed source separation (multisource audio compression), compressed sensing based audio recovery, etc.

Our algorithm is capable of audio signal recovery and source separation in various scenarious, some of which has never been solved by any other algorithm before. We have also extended this algorithm for multichannel problems.

Relevant Publications and Patents

  • Bilen, C.; Ozerov, A.; Pérez, P., ”Solving Time Domain Audio Inverse Problems using Nonnegative Tensor Factorization Model”, IEEE Transactions on Signal Processing, (in review)
  • Bilen, C.; Ozerov, A.; Pérez, P., ”Automatic Allocation of NTF Components for User-Guided Audio Source Separation”, ICASSP 2016
  • Ozerov, A.; Bilen, C.; Pérez, P., ”Multichannel Audio Declipping”, ICASSP 2016
  • Bilen, C.; Ozerov, A.; Pérez, P., ”Audio declipping via nonnegative matrix factorization”, WASPAA 2015
  • Bilen, C.; Ozerov, A.; Pérez, P., ”Compressive Sampling-Based Informed Source Separation”, WASPAA 2015
  • Bilen, C.; Ozerov, A.; Pérez, P., ”Joint Audio Inpainting and Source Separation”, LVA/ICA 2015
  • Bilen, Cagdas; Ozerov, Alexey; Perez, Patrick, ``Method and device for encoding multiple audio signals, and method and device for decoding a mixture of multiple audio signals with improved separation", Patent filed (2015)
  • Bilen, Cagdas; Ozerov, Alexey; Perez, Patrick, ``Method for performing audio restauration, and apparatus for performing audio restauration", Patent filed (2015)