Jingwei Liu

A hybrid SVM/DDBHMM decision fusion modeling for robust continuous digital speech recognition

This paper proposes an improved hybrid support vector machine and duration distribution based hidden Markov (SVM/DDBHMM) decision fusion model for robust continuous digital speech recognition. We investigate the probability outputs combination of support vector machine and Gaussian mixture model in pattern recognition (called FSVM),and embed the fusion probability as similarity into the phone state level decision space of our duration distribution based hidden Markov model (DDBHMM) speech recognition system (named FSVM/DDBHMM). The performances of FSVM and FSVM/DDBHMM are demonstrated in Iris database and continuous mandarin digital speech corpus in 4 noise environments (white, volvo, babble and destroyerengine) from NOISEX-92. The experimental results show the effectiveness of FSVM in Iris data, and the improvement of average word error rate reduction of FSVM/DDBHMM from 6% to 20% compared with the DDBHMM baseline at various signal noise ratios (SNRs) from -5dB to 30dB by step of 5dB.

A DTW-based probability model for speaker feature analysis and data mining

This paper is a contribution to probabilistic data mining and pattern recognition. A DTW-based statistical model is proposed to explore the subspace structures of speaker feature space for feature evaluation, dimension reduction and inter-class information discovery in pattern space. We demonstrate its usefulness in isolated digits speaker identification, and the performance of the statistical model is compared with standard DTW recognition rate in the experiment. We argue that the probability model can be taken as data mining tools.

Feature-Based Causal Structure Discovery in Protein and Gene Expression Data with Bayesian Network

Causal structure discovery is an important problem in protein sequences and gene--gene interaction in gene expression data, which will reveal the elementary structure of the protein sequence and the gene--gene interaction by the expression level of genes within the cell. In this paper, we investigate the feature--based causal structure learning methods for protein sequence and gene expression data respectively. Three feature extraction methods are proposed to model casual structure with Bayesian network with Dirichlet distribution in protein sequence data, and a factor analysis based feature extraction method is discussed for gene expression data Bayesian network learning. The Truncated hemoglobin superfamily from SCOP protein database and Princeton colon gene expression data are involved to demonstrate the causal structure of Bayesian network determined by different feature extraction.