Data-Driven Temporal Filtering on Teager Energy Time Trajectory for Robust Speech Recognition

Data-driven temporal filtering technique is integrated into the time trajectory of Teager energy operation （TEO） based feature parameter for improving the robustness of speech recognition system against noise. Three kinds of data-driven temporal filters are investigated for the motivation of alleviating the harmful effects that the environmental factors have on the speech. The filters include： principle component analysis （PCA） based filters, linear discriminant analysis （LDA） based filters and minimum classification error （MCE） based filters. Detailed comparative analysis among these temporal filtering approaches applied in Teager energy domain is presented. It is shown that while all of them can improve the recognition performance of the original TEO based feature parameter in adverse environment, MCE based temporal filtering can provide the lowest error rate as SNR decreases than any other algorithms.

Robust Speech Recognition Method Based on Discriminative Environment Feature Extraction

It is an effective approach to learn the influence of environmental parameters, such as additive noise and channel distortions, from training data for robust speech recognition. Most of the previous methods are based on maximum likelihood estimation criterion. However, these methods do not lead to a minimum error rate result. In this paper, a novel discrimina-tive learning method of environmental parameters, which is based on Minimum Classification Error (MCE) criterion, is proposed. In the method, a simple classifier and the Generalized Probabilistic Descent (GPD) algorithm are adopted to iteratively learn the environmental pa-rameters. Consequently, the clean speech features are estimated from the noisy speech features with the estimated environmental parameters, and then the estimations of clean speech features are utilized in the back-end HMM classifier. Experiments show that the best error rate reduction of 32.1% is obtained, tested on a task of 18 isolated confusion Korean words, relative to a conventional HMM system.

Stream Weight Training Based on MCE for Audio-Visual LVCSR

In this paper we address the problem of audio-visual speech recognition in the framework of the multi-stream hidden Markov model. Stream weight training based on minimum classification error criterion is discussed for use in large vocabulary continuous speech recognition (LVCSR). We present the lattice re- scoring and Viterbi approaches for calculating the loss function of continuous speech. The experimental re- sults show that in the case of clean audio, the system performance can be improved by 36.1% in relative word error rate reduction when using state-based stream weights trained by a Viterbi approach, compared to an audio only speech recognition system. Further experimental results demonstrate that our audio-visual LVCSR system provides significant enhancement of robustness in noisy environments.

Center-Distance Continuous Probability Models and the Distance Measure

In this paper, a new statistic model named Center-Distance Continuous Probability Model (CDCPM) for speech recognition is described, which is based on Center-Distance Normal (CDN) distribution. In a CDCPM, the probability transition matrix is omitted, and the observation probability density function (PDF) in each state is in the form of embedded multiple-model (EMM) based on the Nearest Neighbour rule. The experimental results on two giant real-world Chinese speech databases and a real-world continuous-manner 2000 phrase system show that this model is a powerful one. Also,a distance measure for CDCPMs is proposed which is based on the Bayesian minimum classification error (MCE) discrimination.