A new codebook design schema for VQ-based Monaural Speech-Music Segregation

Several ideas have been introduced to improve monaural speechmusic segregation problem. Schemadriven approaches employ some statistical methods to model the underlying source signals. Although schemabased techniques present a high quality segregated speech and music outputs, the computational complexity is the main drawback of these methods. In this paper, we proposed an optimized version of hybrid PCA-VQ model based on K-means clustering to overcome this deficiency. K-means algorithm does not work well for high dimensional data in terms of computational complexity and curse of dimensionality issues. To overcome the computational complexity of K-means algorithm and obtain uncorrelated and the most descriptive variables, we used Principal Component Analysis (PCA) technique. This technique is a commonly used statistical approach for dimension reduction. The goal of PCA is a linear mapping which maps data to a lower dimensional space, so that variance of the data in new space is maximized. First, we employed PCA on 2-D framed STFT of the input signal to compute uncorrelated feature vectors efficiently and subsequently. Then we introduced and evaluate a modified version of kmeans method for clustering. The simulation results demonstrate that the proposed schema can improve the segregated procedure in terms of PESQ criterion and complexity measures.