Unsupervised discovery of phoneme boundaries in multi-speaker continuous speech

Abstract

Children rapidly learn the inventory of phonemes used in their native tongues. Computational approaches to learning phoneme boundaries from speech data do not yet reach the level of human performance. We present an algorithm that operates on, qualitatively, similar data to those children receive: natural language utterances from multiple speakers. Our algorithm is unsupervised and discovers phoneme boundary positions in speech. The approach draws inspiration from the word and text segmentation literature. To demonstrate the efficacy of our algorithm on speech data, we present empirical results of our method using the TIMIT data set. Our method achieves F-measure scores in the 0.68 – 0.73 range for locating phoneme boundary positions.