Cochannel Speech Segregation with Sparse Coding

Most of the computational auditory scene analysis (CASA) based systems rely on pitch based features. When we go for cochannel speech segregation, two speakers are involved. Pitch ranges for male speech and female speech overlap to a large extent. Therefore multi-pitch tracking becomes a nontrivial task. In case of same gender mixtures, again pitch tracking becomes harder. Considering this fact, we should go for some reliable features. Here we propose a cochannel speech segregation system with sparsity based features. Sparse coding is applied on the cochleagram of the signal to get sparse approximation coefficients using pre-trained dictionaries for speakers. We treat sparse approximation coefficients the features because these are selected from the speaker specific dictionaries to represent an input signal. Sparse approximation coefficients are good choice for finding binary masks. Speech waveform is resynthesized from the masked cochleagram of the mixture. Experimental results show that the proposed method produces better objective intelligibility scores than the baseline system.