Speech Text-Independent Segmentation Using an Improvement Method for Identification of Phoneme Boundaries

Abstract

The determination of right boundaries during phoneme segmentation of a speech signal is an important part in the process of automatic speech recognition. However, when no information is provided about the meaning of the signal, this segmentation process becomes very difficult. Currently, most of the methods used to detect boundaries of phonemes are based in the identification of variations in distances calculated over a set of features, which are obtained from segments of the signal. Here we present a modification of a previous work, that is based on a different calculation of the distances and a modification in the selection of a boundary. The proposed modification showed to improve the correct segmentation percentage when compared with the previous work, tested in Spanish and English corpus. The improved method obtained 82.59% of correct segmentation over Spanish data, and 80.28% over English data. In addition, the proposed method obtained at average an over-segmentation of 0%.