On the Choice of Window Size in Model-Based Single Channel Speech Separation

In this paper, we study the effect of window size on the performance of model-based single channel speech separation techniques. The separation system consists of two trained codebooks of the log spectral vectors of each speaker where by the main process of separation is performed. Four Hamming windows with the durations of 20, 32, 64, and 128 msec are used for short-time processing of the data. The experimental results show, in contrast with other speech processing applications, that a longer window must be applied for the separation purpose. Improvements in SNR of 2-3 dB are reported by using an optimal window selection which can significantly improve the performance of current model-based speech separation techniques