This paper proposes an improved sound segregation model based on auditory scene analysis in order to overcome three disadvantages in our previous model. This improved model solves the problem of segregating two acoustic sources by using constraints related to the heuristic regularities proposed by Bregman. We first reconsider the estimation method of Ck,0(t) and Dk,0(t), then incorporate the constraint of channel envelopes with periodicity of the fundamental frequency into the grouping block, and finally consider the constraint of smoothness of Ak(t) on channels.
We demonstrated that the proposed model can improve the previous model and that it can precisely segregate real speech from noisy speech in three simulations of segregating two acoustic sources. The evaluations showed that the proposed model can improve the previous model, and that all constraints related to the four regularities are useful in order to segregate the desired vowel from a noisy vowel. Furthermore, the proposed method can precisely segregate the desired signal from noisy signal, compared with basic spectral subtraction.
In the future work, we will (1) do hearing tests for vowel segregation by using the proposed model and some other model, and (2) improve the proposed model so that it can be applied to consonants-vowel segregation.