Segmentation on Time-Frequency Domain for Speech Segregation

In this paper, we propose an algorithm for the frequency channel segmentation using a neural oscillatory network. The frequency channel segments means that local groups of channels in frequency domain that could be arisen from the same sound source. The proposed algorithm is based on the smoothed spectrum of the input sound. Valleys in the smoothed spectrum are used to determine vertical weights and the continuity of segment boundaries is used to determine vertical weights in the oscillatory network. To evaluate a suitableness of the proposed segmentation algorithm before the grouping stage is applied, we compare the synthesis results of ideal mask with that of proposed algorithm