Integrating binaural cues and blind source separation method for separating reverberant speech mixtures

Abstract

This paper presents a new method for reverberant speech separation, based on the combination of binaural cues and blind source separation (BSS) for the automatic classification of the time-frequency (T-F) units of the speech mixture spectrogram. The main idea is to model interaural phase difference, interaural level difference and frequency bin-wise mixing vectors by Gaussian mixture models for each source and then evaluate that model at each T-F point and assign the units with high probability to that source. The model parameters and the assigned regions are refined iteratively using the Expectation-Maximization (EM) algorithm. The proposed method also addresses the permutation problem of the frequency domain BSS by initializing the mixing vectors for each frequency channel. The EM algorithm starts with binaural cues and after a few iterations the estimated probabilistic mask is used to initialize and re-estimate the mixing vector model parameters. We performed experiments on speech mixtures, and showed an average of about 0.8 dB improvement in signal-to-distortion (SDR) over the binauralonly baseline.Index Terms-EM algorithm, interaural phase difference, interaural level difference, blind source separation, mixing vectors

Underdetermined Model-Based Blind Source Separation of Reverberant Speech Mixtures using Spatial Cues in a Variational Bayesian Framework

“…We selected 15 utterances spoken by both male and female speakers at random of the same length (about 3s) and then shortened to 2.5 s as in [5].…”

“…Previous work [1,3,5,7] suggested that by normalizing and pre-whitening the observation the results will be improved. Therefore we will normalize the amplitude of the observations as follows:…”

“…In the mixture model, as in [5], we combine three cues, the T-F observations, the ILD values and the IPD values, denoted by x(ω,t), α(ω,t) and φ (ω,t), respectively, where the total number of time frames is T and the total number of frequency channels is Ω. The total number of initial sources is denoted by I.…”

“…We examined a similar approach by integrating the spatial cues in the separation process. In this paper we model the T-F point as a mixture of complex-Gaussian distribution, similar to [7] and we also model the IPD and ILD as a mixture of Gaussians [4,5]. In this Bayesian framework, we establish proper conjugate prior distributions on the parameters of the model.…”

“…Hence, it is natural to employ the spatial cues in the separation process, such as the interaural level difference (ILD), the interaural phase difference (IPD), and the mixing vectors as examined in [3,4,5], where separation is achieved by modeling the various observations as Gaussian mixtures and applying the Expectation-Maximization (EM) algorithm to obtain the model parameters. However, there are some limitations with the EM algorithm.…”