A hybrid model for unsupervised single channel speech separation

The performance of any voice recognition platform in real environment depends on how well the desired speech signal is separated from unwanted signals like background noise or background speakers. In this paper, we propose a three stage hybrid model to separate two speakers from single channel speech mixture under unsupervised condition. Proposed method combines three techniques namely speech segmentation, NMF (Nonnegative Matrix Factorization) and Masking. Speech segmentation groups the short speech frames belonging to individual speakers by identifying the speaker change over points. The segmentation block groups the speech frames belonging to individual speakers but lacks in continuity of the speech samples. Therefore a second stage is built using NMF. NMF algorithm performs better in separating the speech mixture when parts of the individual speech signals are known a priori. This requirement is satisfied by speech segmentation stage. NMF further separates the individual speech signals in the mixture by maintaining continuity of speech samples over time. To further improve the accuracy of separated speech signals, various masking methods like TFR (Time frequency Ratio), SM (Soft Mask) and HM (Hard Mask) are applied. The separation results are compared with other unsupervised algorithms. The proposed hybrid model produces promising results in unsupervised single channel speech separation. This model can be applied at the front end of any voice recognition platform to further improve the recognition efficiency.