Single channel speech/music segregation based on a novel K-means clustering schema

In this paper, we proposed a modified version of K-means clustering algorithm for single channel separation of speech and music from mixed signal. K-means method fails for high dimensional data processing due to computational complexity and curse of dimensionality issues. To improve the performance of clustering algorithm, we used PCA technique and suggested a novel schema to increase the quality of outcome signals of PCA-Kmeans approach in both FFT and STFT domains. The efficiency of the proposed method is evaluated for different codebook sizes. The comparison between modified PCA-Kmeans algorithm and PCA-Kmeans approach for codebook size 512, showed that the quality of separation signals was improved about 12% in FFT and 20% in STFT without increase in the computational complexity. In addition, the modified PCA-Kmeans algorithm reduced the separation time up to 80% in FFT domain and 85% in STFT domain and improved the quality of segregated speech by about 20% in FFT and STFT domains in comparison with standard K-means method.