Speaker-independent Model-based Single Channel Speech Separation

摘要

In this paper, we present a model-based single channel speech separation (SCSS) technique with two attributes. First, the proposed techniques is speaker-independent. Second, the proposed technique is able to separate out speech signals even though they have been mixed with different levels of energy. A mathematical model is derived in which the probability density function (PDF) of the mixed signal is expressed in terms of envelopes and excitation signals of sources and associated gains. Then a maximum likelihood estimator is used to estimate the sources' parameters and gains. The proposed technique is evaluated with male + male, male + female, and female + female mixtures. The experimental results show a significant signal-to-noise ratio (SNR) improvement when the proposed technique is compared with approaches which apply the excitation signals or log spectra to separate the speech signals in the speaker-independent speech separation scenario.