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