Improved comb filter for the separation of the voiced speech signals in the case of two speakers

In this paper, we present a method for separating voiced sounds from a composite signal. This method is mainly based on the separation by modified comb filter. This filter is keyed to the average values of the estimated pitch. This estimation is performed through an autocorrelation of multi-scale product analysis to separate the effects of the source and the vocal tract. The “autocorrelation of the multi-scale product” method allows noise elimination and the apparition of a signal periodic structure. Peaks that appear are used to calculate the mean fundamental frequency of the target speaker which will be used in the corresponding comb filters to determine the target speaker contribution. After the subtraction of this contribution from the mixture, we obtain the intrusion speaker. This separation is validated by its application on Cooke database and a part of VCTK database and compared to recent methods as Wang–Brown, Hu–Wang, Zhang–Liu and Quan systems. Results confirm the performance of the proposed approach.