Speech Recognition Under Noise Conditions: Compensation Methods

In most of the practical applications of Automatic Speech Recognition (ASR), the input speech is contaminated by a background noise. This strongly degrades the performance of speech recognizers (Gong, 1995; Cole et al., 1995; Torre et al., 2000). The reduction of the accuracy could make unpractical the use of ASR technology in applications that must work in real conditions, where the input speech is usually affected by noise. For this reason, robust speech recognition has become an important focus area of speech research (Cole et al., 1995). Noise has two main effects over the speech representation: it introduces a distortion in the representation space, and it also causes a loss of information, due to its random nature. The distortion of the representation space due to the noise causes a mismatch between the training (clean) and recognition (noisy) conditions. The acoustic models, trained with speech acquired under clean conditions do not model speech acquired under noisy conditions accurately and this degrades the performance of speech recognizers. Most of the methods for robust speech recognition are mainly concerned with the reduction of this mismatch. On the other hand, the information loss caused by noise introduces a degradation even in the case of an optimal mismatch compensation. In this chapter we analyze the problem of speech recognition under noise conditions. Firstly, we study the effect of the noise over the speech representation and over the recognizer performance. Secondly, we consider two categories of methods for compensating the effect of noise over the speech representation. The first one performs a model-based compensation formulated in a statistical framework. The second one considers the main effect of the noise as a transformation of the representation space and compensates the effect of the noise by applying the inverse transformation.