A Blind Method for Phone Segmentation and Its Evaluation on Vietnamese Speech Corpus

A blind method is proposed for phone boundary detection without prior knowledge of number and sequence of phones. Fast Fourier transform is applied on speech signal then the spectrum is filtered to get band energy (BE) feature. BE change is calculated to present spectral variation between speech frames. Then, BE change is increased and decreased by using sigmoid function and quadratic function. The peaks on this curve present locations of possible phone boundaries. To remove spurious peaks, a post processing is used using speech energy feature. Evaluation on English corpus gives us promised results. A brief of Vietnamese phonetics is presented. Good results of the evaluation on Vietnamese corpus indicated that blind method is language independent. However, the performance degrades. Some reasons of degradation are analyzed.