Spectral and textural features for automatic classification of fricatives

Classification of unvoiced fricatives is an important stage in applications such as spoken term detection and audio-video synchronization, and in technologies for the hearing impaired. Due to their acoustic similarity, extraction of multiple features and construction of high-dimensional feature vectors are required for successful classification of these phonemes. In this study two dimensionality reduction algorithms, namely, t-distributed Stochastic Neighbor Embedding (t-SNE) and Sequential Forward Floating Selection (SFFS) were used to obtain a compact representation of the data. A classification stage (kNN or SVM) was then applied, in which we compared the identification rates between the original feature vector and the low-dimensional representation. A total of 1000 unvoiced fricatives (/s/ /sh/ /f/ and /th/) derived from the TIMIT speech database, containing 25000 short frames of 8 ms each, were used for the evaluation. We show that representing the data by a feature vector with as few as 3 dimensions, yields a classification rate of almost 90% which outperforms most of the results obtained in previous studies.

II. Introduction

Subjects with hearing loss in the high frequency range (above 1kHz) suffer from difficulties in discriminating between unvoiced fricatives [1], [2], and in reduced ability to communicate. The audibility of fricatives could be artificially increased using techniques such as amplification of high frequencies, frequency compression or transposition [3], and thus speech perception could be enhanced. Automatic classification of fricatives in conversational speech, is an essential stage in such technologies to allow differential manipulations of these phonemes to improve their discrimination [4].