Temporal filter design by minimum KL divergence criterion for robust speech recognition

Abstract

In this paper, we propose a new temporal filter design method based on minimum KL divergence criterion for robust recognition of noisy and reverberant speech. The main idea is to optimize the filter parameters by minimizing the KL divergence of two distributions, of which one is the feature distribution in the test environment, and another is the feature distribution represented by the acoustic model. The minimization of the KL divergence reduces the mismatch between the acoustic model and the test data. Experimental results on Aurora-5 task shows that the new filter design outperforms other filter design methods significantly in noisy and reverberant test conditions. In addition, the proposed filtering of feature trajectories is shown to be complementary to linear transformation of feature vectors, which is popular in feature processing.

Generalization of temporal filter and linear transformation for robust speech recognition

“…In temporal filtering, if we are using a finite impulse response (FIR) filter such as in TSN [22] and MNLLF [26], we have In linear transformation such as fMLLR [10,14], we have…”

“…Hence, linear transformation uses inter-dimensional correlation information, while temporal filter uses inter-frame correlation information. In our previous study, we have shown that linear transformation and temporal filter are complementary and applying fMLLR after MNLLF [26] produces better results than the two techniques alone.…”

“…One is the linear transformation of feature vectors, such as the feature space maximum likelihood linear regression (fMLLR, also called constrained MLLR) and its extensions [10,[14][15][16][17]. The other is the linear filtering of feature trajectories, such as RASTA [18], ARMA filter [19], data-driven filters [20,21], temporal structure normalization (TSN) and related methods [22][23][24][25], and maximum normalized likelihood linear filter (MNLLF) [26]. From a linear regression point of view, linear transformation uses all dimensions of the current frame to predict the new features that fit the acoustic model under maximum likelihood (ML) criterion.…”

Speech dereverberation for enhancement and recognition using dynamic features constrained deep neural networks and feature adaptation

“…Linear transformation uses all dimensions of the current frame to predict new features that fit the acoustic model under the maximum-likelihood (ML) criterion [30,41]. On the other hand, temporal filtering uses the context information in neighboring frames to estimate features that fit the acoustic model [42][43][44][45]. While linear transformation uses inter-dimensional correlation information (or spectral information) to process features, temporal filtering uses inter-frame correlation information (or temporal information).…”

“…Similar to maximum normalized likelihood linear filtering in [45], the parameters of the cross transform can be estimated by minimizing an approximated KL divergence between the distribution of the processed features, p y , and the distribution of clean training features, p . In this work, p y is modelled by a single Gaussian and p by a GMM with parameter set = {c j , μ j , j |j = 1, .…”