Unsupervised Phoneme Segmentation Using Transformed Cepstrum Features

One of the basic problems in speech engineering is phoneme segmentation, that is, to divide a speech stream into a string of phonemes. Automatic Speech Recognition (ASR) models often require reliable phoneme segmentation in the initial training phase, and Text-to-Speech (TTS) systems need a large speech database with correct phoneme segmentation information for improving the performance. Human speech is a smoothly changing continuous signal. Unlike written language, speech signals don’t include explicit marks, such as space, for segmentation. Moreover, there usually does not exist abrupt changes in speech signals due to the temporal constraints of vocal tract motions. The difficulty of phoneme segmentation comes from co-articulation of speech sounds, where acoustic realization of one phoneme may blend or fuse with its adjacent sounds. This phenomenon can even exist at a distance of two or more phonemes. All these facts make automatic phoneme segmentation a challenging problem. Previous approaches to phoneme segmentation can be divided into two categories: supervised and unsupervised segmentation. In the first case, both the linguistic contents and the the acoustic models of phonemes are available. Thus the segmentation problem can be reduced to align speech signals with a string of acoustic models. Perhaps the most famous approach of this category is HMM-based force alignment [2]. The second category of method tries to perform phonetic segmentation without using any prior knowledge on linguistic contents and acoustic models. The approach of this paper belongs to the 2nd class. The unsupervised segmentation is similar to the situation that infants acquire speech [11]. Infants don’t have acoustic and linguistic models for segmentation. However, psychological facts indicate that infants become able to segment speech according to acoustic difference between speech sounds and cluster speech segments into categories [8]. It is only by this procedure that infants can gradually construct the speech model of their native languages. Most of the previous approaches to this problem focus on detecting the change points of speech