Automatic parameterisation of the glottal waveform combining time and frequency domain measures

Abstract

This paper describes a new technique for automatically parameterising the inverse filtered speech waveform by exploiting frequency domain measures and amplitude measures in the time domain. The technique is motivated by the difficulties posed by time domain analysis and by the consequent risks of inconsistencies on the part of both researchers and time based algorithms. The results demonstrate that the system can obtain accurate measurements on synthetic source signals. Analysis was also carried out on short utterances of three male speakers producing tense, modal and breathy voice qualities. Perception tests which involved comparing different resynthesised utterances provide evidence that the new technique is at least as good as our manual method for modal and tense voices. For breathy voice qualities, however, the system needs further development to include aspects like the noise component to provide a more breathy percept.