Particle filtering based pitch sequence correction for monaural speech segregation

This article proposes a novel speech and sound segregation framework incorporating a technique for correcting a series of pitch periods based on particle filtering. The conventional pitch track correction method finds the peak locations of the autocorrelation functions to estimate the pitch period, and only the longest reliable pitch streak is used to correct unreliable pitch tracks. Especially in noisy environments, it is hard to find long and reliable pitch streaks, resulting in the degradation of the speech segregation performance. The proposed algorithm based on particle filtering considers all the reliable pitch streaks rather than the longest one and smoothly connects the scattered pitch streaks. To apply the particle filtering algorithm to pitch track correction, the importance weight computation to account for the degree of matchness of the found pitch to the individual spectro-temporal components is also proposed. The performance of the proposed method is evaluated by the results of speech segregation experiments for the mixtures of speech and various noise sources in various mixing signal-to-noise ratios (SNRs). The evaluation measures were SNR, energy loss ratio, and noise residue ratio of the segregated speech, and all these measures showed that the proposed segregation method achieved superior performance compared to the conventional approach.