Quantifying parameters of a source-filter model for oesophageal speech

Abstract

Signal processing methods can improve the quality and intelligibility of oesophageal speech. Current methods show only moderate improvement leaving potential for better results. Quantifying parameters of oesophageal speech relative to laryngeal (normal) speech would help in the design of future enhancement methods for oesophageal speech. We quantified parameters of a source-filter model on a database of sustained vowels in Spanish from 4 oesophageal and 4 normal speakers. A ten-parameter glottal waveform model was used as the source and an autoregressive model was used as the filter. Classification, using a log-spectral distance measure, showed that all oesophageal speech samples were classified as whisper voice types; a voice type with a signal to noise ratio of -20 dB. Filter parameters representing spectral amplitudes and bandwidths had a large degree of variation for oesophageal speech comparative to the degree of variation for normal speech (Brown-Forsythe test, F < 0.001). Source metrics, noise to harmonic ratio (NHR) and variation in fundamental frequency, were also significantly greater for oesophageal speech (t-test, P < 0.001). These results show a greater degree of nonstationarity, and a noisier glottal waveform, for oesophageal speech comparative to normal speech.