Data-Driven Filter-Bank-based Feature Extraction for Speech Recognition
Young-Joo Suh, Hoirin Kim · 2004
Abstract Selecting good feature is especially important to achieve high speech recognition accuracy. Although the mel-cepstrum is a popular and effective feature for speech recognition, it is still unclear that the filter-bank in the mel-cepstrum is always optimal regardless of speech recognition environments or the characteristics of specific speech data. In this paper, we focus on the data-driven filter-bank optimization for a new feature extraction where we use the Kullback-Leibler (KL) distance as the measure in the filter-bank design. Experimental results showed that the proposed feature provides an error rate reduction of about 20% for clean speech as well as noisy speech compared to the conventional mel-cepstral feature. 1. Introduction Speech recognition is mainly composed of feature extraction and classification. Of the two parts, feature extraction aims at not only preserving necessary information to distinguish the proper phonetic class from the categorized phonetic ones but also alleviating irrelevant redundancies such as speaker variability, channel variability, or environmental noise [1], [2]. These roles of feature extraction make selection of feature especially important to achieve high recognition accuracy. Currently, most speech recognizers utilize the mel-cepstrum as their input feature because of its predominant attractiveness in speech recognition accuracy as well as noise immunity. The mel-cepstrum is based on the properties of speech production and speech perception, which are reflected by cepstral analysis and the critical band-based filter-bank analysis respectively [3], [4]. Therefore, one of the basic ideas of the mel-cepstrum is to reflect the human auditory perception mechanism on the feature for speech recognition. The relatively superior effectiveness of the mel-cepstral feature to other features is well known from numerous experimental results [3]. Nevertheless, it is still unclear that the filter-bank in the mel-cepstrum is always optimal in the sense of information preservation or speech recognition accuracy regardless of speech recognition environments or the characteristic of speech database for developing a speech recognizer for a specific application domain. This is due to the fact that the mel-scaled filter-bank is mainly based on the results from empirical researches on the human auditory perception [4]. Thus, there are always some needs to make new approaches in the sense of maximizing the preservation of information driven from real speech data in the feature extraction. As a part of solving these topics, several research activities have been conducted to optimize the filter-bank of the mel-cepstrum [1], [2] and [5]. Our work, as another approach to this research area, focuses on the filter-bank optimization for a new feature extraction for the given speech data environments. Here, we use an optimization criterion derived from the information theory-based entropic distance measure in the process of filter-bank design. This paper organized as follows. Section 2 describes the overall algorithmic procedure used for designing the optimized filter-bank for the newly proposed feature extraction. Section 3 presents experimental results for evaluating the performance of the proposed feature extraction method. Conclusion is finally given in Section 4.