New Distance Measure for Monaural Model-based Sound Separation

We present a new distance measure based on certain perceptual cues to improve the state-of-the-art model-based sound separation performance which has been introduced as a challenging topic for decades. Conducting several simulation results, it is demonstrated that using such an appropriate distance measure proposed in this work in place of the commonly used Euclidean distance in Vector Quantization (VQ) procedure can significantly result in a better speech modeling which is also close to transparent reconstruction in terms of naturalness. It is also observed that choosing an overlap around 25terms of both Spectral Distortion Ratio (SDR) and Segmental Signal to Noise Ratio (SSNR).