Study of Associative Cepstral Statistics Normalization Techniques for Robust Speech Recognition in Additive Noise Environments

Computational Linguistics and Chinese Language Processing Vol. 14, No. 1, March 2009, pp. 105-132 105 © The Association for Computational Linguistics and Chinese Language Processing [Received January 13, 2009; Revised May 13, 2009; Accepted May 19, 2009] Study of Associative Cepstral Statistics Normalization Techniques for Robust Speech Recognition in Additive Noise Environments Wen-Hsiang Tu∗ and Jeih-weih Hung∗ Abstract Feature statistics normalization techniques have been shown to be very successful in improving the noise robustness of a speech recognition system. In this paper, we propose an associative scheme in order to obtain a more accurate estimate of the statistical information in these techniques. By properly integrating codebook and utterance knowledge, the resulting associative cepstral mean subtraction (A-CMS), associative cepstral mean and variance normalization (A-CMVN), and associative histogram equalization (A-HEQ) behave significantly better than the conventional utterance-based and codebook-based versions in additive noise environments. For the Aurora-2 clean-condition training task, the new proposed associative histogram equalization (A-HEQ) provides an average recognition accuracy of 90.69%, which is better than utterance-based HEQ (87.67%) and codebook-based HEQ (86.00%). Keywords: Speech Recognition, Noise-Robust Feature, Codebook 1. Introduction The performance of a speech recognition system is often severely degraded when there is a mismatch between the acoustic conditions of the training and the application environments. This mismatch may come from various sources, such as additive noise, channel distortion, different speaker characteristics, and different speaking modes. A variety of robustness techniques with demonstrated improvement in system performance have been proposed to reduce this mismatch. For the purpose of handling additive noise, these robustness techniques can be roughly divided into three classes: adaptation of the speech models in the recognizer to make them better match the noise conditions, enhancement of the speech features before they are fed to the recognizer, and utilization of a noise robust representation of speech signals. In ∗ Dept of Electrical Engineering, National Chi Nan University, Nantou County, Taiwan, Republic of China E-mail: aero3016@ms45.hinet.net; jwhung@ncnu.edu.tw 106 Wen-Hsiang Tu and Jeih-weih Hung the first class of approaches, compensation is performed on the pre-trained recognition model parameters so that the modified recognition models can more effectively classify the mismatched testing speech features collected in the application environment. Typical examples of this class include the well-known noise masking (Holmes & Sedgwick, 1986; Klatt, 1979; Nadas, Nahamoo, & Picheny, 1988), speech and noise decomposition (SND) (Varga & Moore, 1990), hypothesized Wiener filtering (Berstein & Shallom, 1991; Beattie & Young, 1992), vector Taylor series (VTS) (Acero, Deng, Kristjansson, & Zhang, 2000), maximum likelihood linear regression (MLLR) (Leggester & Woodland, 1995), model-based stochastic matching (Sankar & Lee, 1996; Lee, 1998), statistical re-estimation (STAR) (Moreno, Raj, & Stem, 1996), and parallel model combination (PMC) (Gales & Young, 1993; 1995a; 1995b). In the second class of approaches, the obtained testing speech features are modified in order to fit the acoustic conditions of pre-trained recognition models more compatibly. Examples of this class include the well-known spectral subtraction (SS) (Boll, 1979), codeword-dependent cepstral normalization (CDCN) (Acero, 1990), feature-based stochastic matching (Sankar & Lee, 1996; Lee, …

Content truncated to 500 words