Speech Dereverberation Based on Variance-Normalized Delayed Linear Prediction

Probabilistic integration of diffuse noise suppression and dereverberation

“…As we see in (1), in this paper, we omit the subscript f for brevity, which should not cause confusion, because the proposed methods process each frequency bin independently. For simplicity, we assume that xt and vt are Gaussian random variables with zero mean and time-varying covariance matrices Φ…”

“…That is, we aim to estimate the dereverberated desired signal from noisy, reverberant observations. To this end, we extend the probabilistic model for denoising in Section 2.1 by combining it with a probabilistic model for dereverberation [1,4]. In Section 2.1, we modeled xt as temporally independent.…”

“…Linear prediction based dereverberation [1][2][3][4][5] is among the most promising dereverberation techniques [6][7][8]. It exploits quasistationarity of the desired speech and predictability of reverberation as the above-mentioned structures.…”

“…It exploits quasistationarity of the desired speech and predictability of reverberation as the above-mentioned structures. On the one hand, the desired speech is correlated only within several tens of milliseconds, and approximately uncorrelated for a longer time lag [1]. This structure is reflected in speech modeling as a time-varying random variable that is uncorrelated across time frames.…”

“…On the other hand, reverberation is predictable using past samples. This structure is modeled as an autoregressive model [1], which excludes a few most recent samples from prediction to avoid whiting out the above short-term correlation of speech. These models enable distinction between the desired speech and reverberation and estimation of prediction gains for dereverberation through likelihood maximization.…”