Speech Recognition in Unknown Noisy Conditions

Abstract

This chapter describes our recent advances in automatic speech recognition, with a focus on improving the robustness against environmental noise. In particular, we investigate a new approach for performing recognition using noisy speech samples without assuming prior information about the noise. The research is motivated in part by the increasing deployment of speech recognition technologies on handheld devices or the Internet. Due to the mobile nature of such systems, the acoustic environments and hence the noise sources can be highly time-varying and potentially unknown. This raises the requirement for noise robustness in the absence of information about the noise. Traditional approaches for noisy speech recognition include noise filtering or noise compensation. Noise filtering aims to remove the noise from the speech signal. Typical techniques include spectral subtraction (Boll, 1979), Wiener filtering (Macho et al., 2002) and RASTA filtering (Hermansky & Morgan, 1994), each assuming a priori knowledge of the noise spectra. Noise compensation aims to construct a new acoustic model to match the noisy environment thereby reducing the mismatch between the training and testing data. Typical approaches include parallel model combination (PMC) (Gales & Young, 1993), multicondition training (Lippmann et al., 1987; Pearce & Hirsch, 2000), and SPLICE (Deng et al., 2001). PMC composes a noisy acoustic model from a clean model by incorporating a statistical model of the noise; multicondition training constructs acoustic models suitable for a number of noisy environments through the use of training data from each of the environments; SPLICE improves noise robustness by assuming that stereo training data exist for estimating the corruption characteristics. More recent studies are focused on the approaches requiring less information about the noise, since this information can be difficult to obtain in mobile environments subject to time-varying, unpredictable noise. For example, recent studies on missing-feature theory suggest that, when knowledge of the noise is insufficient for cleaning up the speech features, one may alternatively ignore the severely corrupted features and focus the recognition only on the features with little or no contamination. This can effectively reduce the influence of noise while requiring less knowledge than usually needed for noise filtering or compensation (e.g., Lippmann & Carlson, 1997; Raj et al., 1998; Cooke et al., 2001; Ming et al., 2002). However, missing-feature theory is only effective given partial feature corruption, i.e., the noise only affects part of the speech representation and the remaining part not