Casa Based Supervised Single Channel Speaker Independent Speech Separation

Abstract

Computational auditory scene analysis (CASA) based speech separation is widely considered in a number speech processing applications and is used to separate a target speech from target-interference mixtures and usually the task of target separation is considered as a signal processing problem. However, target speech separation is formulated as a supervised learning problem and discriminative patterns of speech, speakers and background noises are learned from input training data. In this paper, we present a single channel supervised speech separation approach based on the ideal binary mask (IBM) estimation. In proposed approach, speaker independent speech separation system is trained with sets of the clean speech magnitudes and during separation; SNR is estimated in time-frequency (TF) channels using clean magnitudes and compared to a pre-defined threshold. The TF channels satisfying threshold are hold while TF channels violating the threshold are discarded to construct an IBM. The estimated mask is than applied to the mixtures to reconstruct the target speech, using phase of the mixture speech. The experiments are conducted in three speaker independent mixture's scenarios: termed as 2-talkers, 3talkers and 4-talkers mixtures at four input SNRs: -5dB, 0dB, 5dB and 10dB. The experimental outcomes reported that proposed CASA based supervised speaker independent mask estimation outperformed the competing approaches: Nonnegative matrix factorization (NMF), Nonnegative dynamical system (NNDS) and log minimum mean square error (LMMSE) estimation in terms of PESQ, SegSNR, LLR, WSS, SIG, BAK and STOI objective measures.