Glottal Waveform Model for Oesophageal Speech

Oesophageal speech, a mode of speech for laryngectomees, is of low quality and intelligibility comparative to normal (laryngeal) speech. Understanding the signal differences between oesophageal and normal speech will help future oesophageal speech enhancement methods. We aim to produce a method to synthesise oesophageal speech using a simple source–filter model. In this paper, we fit a parametric glottal waveform model to speech samples in our oesophageal database. (This glottal waveform represents the source component in the source–filter approach.) We added coloured noise to the glottal waveform model to produce realistic sounding oesophageal speech. Our fitting error measure, a spectral distance measure, reduces for all tested speech samples when adding the coloured noise. Yet missing from our synthesised signal is the rough-sounding quality often present in oesophageal speech. This work represents the first steps in developing a method to synthesise oesophageal speech.