Effect of the Environment on Speech Statistics

This paper describes a series of cepstral-based compensation procedures that render the SPHINX-II continuous speech recognition system more robust with respect to acoustical changes in the environment. The first two algorithms, SNR based MultivaRiate gAussian based cepsTral normaliZation (SNR-based RATZ) and STAtistical Reestimation of HMMs (STAR), compensate for environmental degradation based on comparisons of simultaneouslyrecorded data in the training and testing environments (“stereo data”). They differ in that RATZ modifies the incoming feature vectors to a recognition system while STAR modifies the internal representation of speech by the system. We also describe N-CDCN, an improved version of codeword-dependent cepstral normalization (CDCN) which does not require stereo training data but nevertheless achieves performance levels comparable to RATZ and other algorithms that require stereo training. Use of these compensation algorithms significantly reduces the error rates for SPHINX-II. The algorithms are tested in a variety of databases and environmental conditions. INTRODUCTION Robustness with respect to environmental variability remains a continuing problem for speech recognition technology (e.g. [3]). For example, the use of microphones other than the ARPA standard Sennheiser HM-414 headset (CLSTLK) severely degrades the performance of speech recognition systems like the SPHINX-II, even in relatively quiet environments [1, 4]. Traditional algorithms to compensate for environmental variation have either relied on the availability of simultaneously-recorded data in the training and testing environments (“stereo data”), or have utilized structural models to define the degradation. For example, multiple fixed codeword-dependent cepstral normalization (MFCDCN) [4] uses stereo data to compute correction vectors to compensate for the effects of the environment. Dual-channel codebook adaptation (DCCA) [4], on the other hand, modifies the statistical representation used in the HMMs for speech on the basis of comparisons obtained from stereo data. The other approach to environmental compensation is through the use of structural models of degradation. On the other hand, codeword-dependant cepstral normalization (CDCN) [1] assumes that speech is degraded by unknown additive noise and unknown linear filtering. It makes use of expectationmaximization (EM) techniques to determine the parameters characterizing these distortions. In this paper we describe three new cepstral-domain compensation strategies, SNR based MultivaRiate gAussian based cepsTral normaliZation (SNR-based RATZ), STAtistical Reestimation of HMMs (STAR), and new CDCN (NCDCN). SNR-based RATZ and STAR both make use of stereo data. They differ in that RATZ modifies the incoming feature vectors to a recognition system while STAR modifies the internal representation of speech by the system. RATZ and STAR are similar in philosophy to MFCDCN and DCCA, respectively [4], but they achieve improved performance through the use of better mathematical models which introduce strong structural constraints into the assumed distribution for speech. N-CDCN is a modification and improvement of the original CDCN algorithm, which is based on a structural model of degradation. EFFECT OF THE ENVIRONMENT ON SPEECH STATISTICS In this section we describe how even well-behaved environments, such as those that can be modeled by unknown linear filtering and additive stationary noise, modify the statistics of “clean” speech in very unpredictable ways. Even though we can formulate equations that analytically describe how the pdfs of clean speech change, the solutions for these equations are mathematically intractable. For analytical purposes, we adopt the simple model of degradation proposed by Acero [1]. In this model, degraded speech is characterized by passing high-quality clean speech through a linear filter and contaminating the filtered output by additive stationary noise. For simplicity, we will also assume that the feature vector is unidimensional, APPROACHES TO ENVIRONMENT COMPENSATION IN AUTOMATIC SPEECH RECOGNITION Pedro J. Moreno, Bhiksha Raj, Richard M. Stern Department of Electrical and Computer Engineering and School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, U.S.A. although all conclusions developed can be easily extended to an arbitrary N-dimensional space such as the log spectral domain. The degraded speech can be characterized as: (1) where represents the power spectrum of the degraded speech, is the power spectrum of the clean speech, is the transfer function of the linear filter, and is the power spectrum of the additive noise. In the logspectral domain this relation can be expressed as: