Detection of burst onset landmarks in speech using rate of change of spectral moments

Burst onset landmarks in the speech signal are transient segments with low energy and their accurate detection is important in applications involving landmark based speech modification, estimation of place of closure for speech training aids, and phoneme recognition. Rate of change measures of energy parameters from spectral bands with fixed boundaries are generally used for landmark detection. The differences in the parameter variation rates and ranges, correlations among them, and their dependencies on speakers, make them less suitable for precise time localization of burst onsets. A method for detection of burst onset landmarks is presented which uses rate of change of spectral moments, in addition to energy parameters of the short-time speech spectrum. Evaluation results indicate that this method can give high detection rates with improved temporal accuracy.