Harmonics Template and Application in Computational Auditory Scene Analysis for Co-channel Speech

In this paper, harmonics template was proposed to analyze harmonics in co-channel speech. New feature first-peak position of autocorrelation was exploited when we generated and utilized harmonics template. We gain harmonics template by statistics and curve fitting. For co-channel speech, harmonics from different channels were achieved by applying frequency channel piecewise continuity and matching with harmonics template. Then, we proposed selected harmonics method to compute the normalized sum of autocorrelation for multi-pitch detection. By comparison with previous algorithms, it was showed that proposed algorithm had better effects in multi-pitch detection. Also, we applied harmonics template to multi-level unit mask estimation in CASA. Although it was not much better than primal 0-1 mask by primary experiment, a good future was shown..