Segmentation of speech signal into phonemes using two-level GMM tokenization

This paper proposes an algorithm for identifying the phoneme boundaries in a given speech signal without the need for its orthographic transcription. The algorithm is a two level process whereby in the first level the phoneme boundaries are determined by silence/voiced/unvoiced classification and in the second level the voiced parts are alone tokenized further. TIMIT database was used to carry out the experiments and to check the correctness of the automatically detected phoneme boundaries. The experimental results showed that the performance of the algorithm in identifying the correct boundaries was ~75%.