Deep Clustering in Complex Domain for Single-Channel Speech Separation
Runling Liu, Yu Tang, Hongwei Zhang · 2022 IEEE 17th Conference on Industrial Electronics and Applications (ICIEA) · 2022
Despite the great success of deep clustering (DPCL) technique in speaker-independent single-channel speech separation has been achieved, phase reconstruction of each source is still a challenging problem. In this paper, a novel method based on DPCL is proposed to avoid this problem, which is named deep clustering in complex domain (DPCL-CD). Different from DPCL, DPCL-CD is trained in complex domain to obtain two embedding vectors for each time-frequency bin. The trained network is thence utilized in testing stage to estimate the binary masks of the real and imaginary parts by clustering. The similarity score is calculated to match the binary masks of the real part with the imaginary part, and then waveforms in time domain can be reconstructed directly without using the phase of mixture. Experiments are carried out to evaluate our method with two-speaker mixed task on the TIMIT and AISHELL-1 dataset, and comparisons on the speech separation performance are made between DPCL and our method.