Bark Wavelet Transform And Its Application For Overlapping Speech Separation

This paper explains the conception of auditory scene analysis (ASA) and its effect in speech signals processing. With Bark wavelet and frequency domain transform, this paper describes an approach to overlapping speech signals separation. It is comprised of overlapping speech analysis stage and waveform synthesis stage. In the former stage, Based on the concept that both glottal excitation and vocal tract parameters are time varying in amplitude and frequency, we detect pitch with frequency domain transform, Bark wavelet transform and (enhanced) autocorrelation algorithm. In the later stage, a series of short-time period signals are partitioned from overlapping speech signals. With the frequency-selection-characteristic of digital multivibrator, a set of amplitudes, frequencies and phases of harmonics based on pitch flow getting from analysis stage act as the parameters of synthesis model, and generate the synthetic speech waveform. The results obtained indicate that the approach described here is efficient.