A novel method for estimating the number of speakers based on generalized eigenvalue-vector decomposition and adaptive wavelet transform by using K-means clustering

Abstract

The aim of this article is estimating the number of simultaneous speakers from the overlapped speech signals. The percentage of correct number of speakers is an important factor for the proposed algorithm. The proposed method in this article is based on spectrum estimation by using the adaptive wavelet transform in combination with generalized eigenvalue-vector decomposition (GEVD) and K-means clustering. Firstly, the speech signals are obtained by a uniform circular array, and each adjacent microphone pairs are considered for the processing. Then, the spectral estimation method is implemented on all microphone signals to select the best part of the speech spectrum. Next, the microphone signals are divided into different subbands by using adaptive wavelet transform. The GEVD algorithm is implemented on each microphone pairs in different subbands and time frames to estimate the room impulse response and time difference of arrival (TDOA). Finally, the K-means clustering with silhouette criteria is used to estimate the number of speakers (K value). The proposed algorithm is implemented on simulated and real data to show the superiority of proposed method in comparison with PENS, Bessel, i-vector PLDA, Hilbert envelope and DNN-based method. The proposed scheme outperforms the other evaluated schemes by 18% in terms of correct estimations in noisy-reverberant conditions for five simultaneous speakers.