Source Separation Based on Non-Negative Matrix Factorization of the Synchrosqueezing Transform

Neha Singh (1) , Sylvain Meignen (1) , Thomas Oberlin (2)

1 DAO - Données, Apprentissage et Optimisation
2 ISAE-SUPAERO - Institut Supérieur de l'Aéronautique et de l'Espace

In this paper, we consider the problem of single channel blind source separation, for which a very common and effective solution consists of applying non-negative matrix factorization to the spectrogram of the mixture. We here propose to replace the spectrogram by the modulus of the synchrosqueezing transform (SST), which achieves a sharper time-frequency representation. Then we introduce two methods for reconstructing the sources, one based on the direct reconstruction from the synchrosqueezed representation, and the other on a twostep procedure based on both the short-time Fourier transform (STFT) and SST, the latter technique being introduced to deal with large signals. Our experiments suggest that non-negative matrix factorization applied to SST enables a better source separation than when applied to the modulus of STFT, and that the proposed two-step procedure using SST and STFT also performs better than the classical technique based on STFT only.

Mots clés: Synchrosqueezing transform; Non negative matrix factorization; Short-time Fourier transform; time-frequency reassignment

Domaines: Recherche d'information [cs.IR]; Analyse fonctionnelle [math.FA]

Neha Singh, Sylvain Meignen, Thomas Oberlin. Source Separation Based on Non-Negative Matrix Factorization of the Synchrosqueezing Transform. 29th European Signal Processing Conference (EUSIPCO), Aug 2021, Dublin, Ireland. ⟨10.23919/EUSIPCO54536.2021.9616194⟩. ⟨hal-03610139⟩