Monaural voiced speech separation with multipitch tracking

Separating voiced speech from its mixtures with interferences in monaural condition is not only an important but also challenging task. As multipitch tracking can enable much better performance of speech separation for CASA systems, we propose a new multipitch determination algorithm, which can be used under various kinds of noise conditions. In the process of multipitch estimation, a new representation method is utilized together with maximum support constraint and harmonic completeness constraint. The proposed approach can reliably detect up to two pitches in each frame. Sequential grouping is performed based on a new target pitch tracking strategy. System evaluations show that our algorithm leads to significantly better speech separation results than previous ones.