Model-Based Monaural Sound Separation by Split-VQ

Abstract—In many speech separation and enhancement tech-niques, establishing a statistical model like a Vector Quantization (VQ) is a must to handle the so-called model-based approaches. It is also desirable to establish a trade-off between sparsity and accuracy in the quantizer. To do so, in this paper we present split-VQ for sinusoidal parameters. We observed that sinusoidal parameters including amplitudes and frequencies, are most capable to be used as our features for split-VQ since they can be easily mapped to a tree-like structure. We demonstrate that using such split-VQ along with fixed dimension sparse sinusoidal parameters can significantly result in better source model compared with common STFT feature vectors in terms of objective and subjective measures in model-based approaches like monaural sound separation