Posterior-Based Features and Distances in Template Matching for Speech Recognition

Abstract

Abstract. The use of large speech corpora in example-based approaches for speech recognition is mainly focused on increasing the number of examples. This strategy presents some difficulties because databases may not provide enough examples for some rare words. In this paper we present a different method to incorporate the information contained in such corpora in these examplebased systems. A multilayer perceptron is trained on these databases to estimate speaker and task-independent phoneme posterior probabilities, which are used as speech features. By reducing the variability of features, fewer examples are needed to properly characterize a word. In this way, performance can be highly improved when limited number of examples is available. Moreover, we also study posterior-based local distances, these result more effective than traditional Euclidean distance. Experiments on Phonebook database support the idea that posterior features with a proper local distance can yield competitive results.

Posterior features applied to speech recognition tasks with user-defined vocabulary

“…This approach is based on TM where speech features are phoneme posterior estimates. In this paper, we confirm the suitability of applying KL divergence when using posterior features as observed in previous experiments [5]. We also show that a weighted combination of the KL divergence can further improve the accuracy.…”

“…previous experiments [5] have shown that the use of the KL divergence can yield better performance when using posterior features. In the next section, we describe the local distances used in this work.…”

“…This approach thus combines the advantages of both TM and HMM-based approaches because it benefits from a simple implementation and a fast decoding time and also, the speech variability from an auxiliary database can be incorporated through the posteriors at the feature level. Moreover, since phoneme posterior features can be seen as discrete probability distributions over the phoneme space, measures coming from the information theory field such as the entropy and the Kullback-Leibler (KL) divergence [4] can be successfully applied [5]. A scheme of this approach is shown in Figure 1(b).…”

An Acoustic Model Based on Kullback-Leibler Divergence for Posterior Features

“…We also show that hybrid HMM/ANN system can be interpreted as a particular case of this model where state target distributions are fixed and equal to a delta distribution. Furthermore, this system naturally extends our previous work where we successfully applied posterior features and KL-divergence to the template matching approach for ASR [8].…”

Learning better lexical properties for recurrent OOV words

“…Acoustic features, such as the melscale frequency cepstral coefficients (MFCCs), are highly sensitive to speaker and channel variations. On the other hand, posteriorbased features, such as the phonetic posteriorgram, are more robust and also widely used in speech recognition [19][20][21]. Therefore, we used the posterior feature to measure the acoustic distance between OOV candidates.…”