Segmentation of a speech waveform according to glottal open and closed phases using an autoregressive-HMM

This paper presents an algorithm to segment speech according to glottal open and closed phases using the time waveform alone. Based on this, pitch, jitter and closed to open glottal ratios can be computed. Segmentation is achieved by identifying spectral changepoints at the subpitch period timescale. Changepoints are identi ed using a 3-state autoregressive hidden Markov model (AR-HMM) operating on the time waveform, with the LiljencrantsFant (LF) glottal model as a theoretical basis. Model parameters and optimal state sequence are determined respectively using the expectation-maximisation (EM) algorithm and a bounded state duration (BSD) Viterbi algorithm. Experiments on synthetic speech give encouraging glottal segmentation for modal, fry and breathy voice types. Experiments on real speech obtained from TIMIT give meaningful segmentations also.