A Pitch State Dependent Dictionary Design Method for Single-Channel Speech Separation

In this paper, we propose to design a new pitch state dependent dictionary to perform single-channel speech separation. The dictionary is designed in two stages, which are sub-dictionary learning and sub-dictionary concatenation. In sub-dictionary learning, pitch state information is taken into account to learn a set of discriminative sub-dictionaries for each speaker in time-domain. To be specific, each sub-dictionary is generated as a matrix composed of the speaker's training frames of similar pitch states as columns. Moreover, we utilize a frequent pattern mining method to further reduce the sub-dictionary size. In sub-dictionary concatenation, we propose to select an appropriate weight pair to match the learned sub-dictionaries to generate a dictionary for separation. Experimental results show that the proposed method achieves better overall performance than two dictionary-based methods and a source-filter-based method also using pitch information.