Deep Clustering With Constant Q Transform For Multi-Talker Single Channel Speech Separation

Deep clustering technique is a state-of-the-art deep learning-based method for multi-talker speaker-independent speech separation. It solves the label ambiguity problem by mapping time-frequency (TF) bins of the mixed spectrogram to an embedding space, and assigning contrastive embedding vectors to different TF regions in order to predict the mask of the target spectrogram of each speaker. The original deep clustering transforms the speech into the TF domain through a short-time Fourier transform (STFT). Since the frequency component of STFT is linear, while the frequency distribution of human auditory system is nonlinear. Therefore, we propose to use constant Q transform (CQT) instead of STFT to achieve a better simulation of the frequency resolving power of the human auditory system. The ideal upper bound of signal-to-distortion (SDR) of CQT based deep clustering is higher than that based on STFT. In the same experimental setting on WSJ0-2mix corpus, we gave a detail description in selecting meta-parameters of CQT for speech separation, and finally the SDR improvements of this method achieved about 1dB better performance than the original deep clustering