Single channel signal separation using pseudo-stereo model and time-freqency masking

Abstract

In many practical applications, one sensor is only available to record a mixture of a number of signals. Single-channel blind signal separation (SCBSS) is the research topic that addresses the problem of recovering the original signals from the observed mixture without (or as little as possible) any prior knowledge of the signals. Given a single mixture, a new pseudo-stereo mixing model is developed. A “pseudo-stereo” mixture is formulated by weighting and time-shifting the original single-channel mixture. This creates an artificial resemblance of a stereo signal given by one location which results in the same time-delay but different attenuation of the source signals. The pseudo-stereo mixing model relaxes the underdetermined ill-conditions associated with monaural source separation and begets the advantage of the relationship of the signals between the readily observed mixture and the pseudo-stereo mixture. This research proposes three novel algorithms based on the pseudo-stereo mixing model and the binary time-frequency (TF) mask. Firstly, the proposed SCBSS algorithm estimates signals’ weighted coefficients from a ratio of the pseudo-stereo mixing model and then constructs a binary maximum likelihood TF masking for separating the observed mixture. Secondly, a mixture in noisy background environment is considered. Thus, a mixture enhancement algorithm has been developed and the proposed SCBSS algorithm is reformulated using an adaptive coefficients estimator. The adaptive coefficients estimator computes the signal characteristics for each time frame. This property is desirable for both speech and audio signals as they are aptly characterized as non-stationary AR processes. Finally, a multiple-time delay (MTD) pseudo-stereo SINGLE CHANNEL SIGNAL SEPARATION ii mixture is developed. The MTD mixture enhances the flexibility as well as the separability over the originally proposed pseudo-stereo mixing model. The separation algorithm of the MTD mixture has also been derived. Additionally, comparison analysis between the MTD mixture and the pseudo-stereo mixture has also been identified. All algorithms have been demonstrated by synthesized and real-audio signals. The performance of source separation has been assessed by measuring the distortion between original source and the estimated one according to the signal-to-distortion (SDR) ratio. Results show that all proposed SCBSS algorithms yield a significantly better separation performance with an average SDR improvement that ranges from 2.4dB to 5dB per source and they are computationally faster over the benchmarked algorithms.