Vq-Based Approach To Single-Channel Audio Separation For Music and Speech Mixtures

In this paper, we propose a low-complexity model-based single-channel audio separation approach. theproposed method presents three certain advantages over previous methods: 1) replacing commonly used linear maskslike wiener filtering by a proposed non-linear one, we show that it is possible to lower the crosstalk of the interferingsource often occurring in a mask-based method while recovering the underlying signals from the observed mixture.using nonlinear masks establishes a tradeoff between acceptable level of interference and low speech distortion, 2) asa post-processing stage, we use phase synchronization technique to enhance the perceptual quality of the resynthesizedsignals, and 3) the proposed method is based on vector quantization (vq) codebooks. hence, thecomplexity is lower than previous gmm-based methods. through extensive experiments, it is demonstrated that theproposed method can achieve a lower signal-to-distortion ratio (sdr). according to our listening experiments andaccording to the mean opinion score (mos) results, it is confirmed that the proposed method is able to recoverseparated outputs with a higher perceived signal quality.