Dimension Reduction and Discriminant Analysis for Japanese Connected Vowel Recognition

The aim of speech recognition is to extract only the linguistic information from speech signals. The acoustic variations caused by non-linguistic factors, such as speaker, communication channel and noise, pose a challenging problem for speech recognition. The same text can lead to different acoustic observations due to different speakers and different environments. To deal with these variations, modern speech recognition approaches mainly make use of the statistical methods (such as GMM, HMM) to model the distributions of the acoustic features. These methods can achieve relatively high recognition rates when properly trained. However, they always require a large number of high quality data for training. This is very different from children spoken language acquisition, where the children mainly use very biased training data from mothers and fathers. This fact largely indicates that there may exist robust representations of speech which are nearly invariant to non-linguistic variations. Along this line, the third author of this paper proposed an invariant structural representation of speech signals, which tries to remove the nonlinguistic factors in speech signals [1]. Different from classical speech models, this structural representations focus on the dynamic motions in speech and discard the static features. Mathematically, the structural representations are made up of Bhattacharyya distances (BD), which are invariant to invertible transformations on feature space [2]. Our previous works have demonstrated the effectiveness and efficiency of this novel representation in both speech recognition tasks [3, 4] and computer aided language learning (CALL) systems [5]. However, there are two limitations for direct use of structural representation for speech recognition. 1) Its dimension is high, which not only increases the computational cost but also makes it easily suffer from the curse of dimensionality [3]. 2) The invariance can be too strong, such that two linguistically different speech signals may have similar structural representations [4]. In this paper, we introduce the techiniques of dimension reduction and discriminant analysis to address these two problems. As first, we build a structure for each sub-stream of the cepstrum features to overcome the too strong invariance. Then we calculate a reduced structure vector for each sub-stream and apply linear discriminant analysis for final classification. The new represen-