Online Speaker Separation Using Deep Clustering

Abstract

In this thesis, a low-latency variant of speaker-independent deep clustering method is proposed for speaker separation. Compared to the offline deep clustering separation system, bidirectional long-short term memory networks (BLSTMs) are replaced with long-short term memory networks (LSTMs). The reason is that the data has to be fed to the BLSTM networks both forward and backward directions. Additionally, the final outputs depend on both directions, which make online processing not possible. Also, 32 ms synthesis window is replaced with 8 ms in order to cooperate with low- latency applications like hearing aids since the algorithmic latency depends upon the length of synthesis window. Furthermore, the beginning of the audio mixture, here, referred as buffer, is used to get the cluster centers for the constituent speakers in the mixture serving as the initialization purpose. Later, those centers are used to assign clusters for the rest of the mixture to achieve speaker separation with the latency of 8 ms. The algorithm is evaluated on the Wall Street Journal corpus (WSJ0). Changing the networks from BLSTM to LSTM while keeping the same window length degrades the separation performance measured by signal-to-distortion ratio (SDR) by 1.0 dB, which implies that the future information is important for the separation. For investigating the effect of window length, keeping the same network structure (LSTM), by changing window length from 32 ms to 8 ms, another 1.1 dB drop in SDR is found. For the low-latency deep clustering speaker separation system, different duration of buffer is studied. It is observed that initially, the separation performance increases as the buffer increases. However, with buffer length of 0.3 s, the separation performance keeps steady even by increasing the buffer. Compared to offline deep clustering separation system, degradation of 2.8 dB in SDR is observed for online system.