Perceptual and Acoustic Impacts of Aberrant Properties of Electrolaryngeal Speech

Advanced laryngeal cancer is often treated by surgical removal of the larynx (laryngectomy) thus rendering patients unable to produce normal voice and speech. Laryngectomy patients must rely on an alternative means of producing voice and speech, with the most common method being the use of an electrolarynx (EL). The EL is a small, hand-held, electromechanical device that acoustically excites the vocal tract when held against the neck or at the lips. While the EL provides a serviceable means of communication, the resulting speech has several shortcomings in terms of both intelligibility and speech quality. Previous studies have identified and tried to correct different single selected acoustic properties associated with the abnormal quality of EL speech, but with only limited success. There remains uncertainty about: 1) which components of the EL speech acoustic signal are contributing most to its abnormal quality and 2) what kinds of acoustic enhancements would be most effective in improving the quality of EL speech. Using a combination of listening experiments, acoustic analysis and acoustic modeling, this thesis investigated the perceptual and acoustic impacts of several aberrant properties of EL speech, with the overall goal of using the results to direct future EL speech improvement efforts. Perceptual experiments conducted by having 10 listeners judge the naturalness of differently enhanced versions of EL speech demonstrated that adding pitch information would produce the most benefit. Removing the EL self-noise and correcting for a lack of low frequency energy would also improve EL speech, but to a lesser extent. However, this study also demonstrated that monotonous, normal speech was found to be more natural than any version of EL speech, indicating that there are other abnormal properties of EL speech contributing to its unnatural quality. An acoustic analysis of a corpus of preand post-laryngectomy speech revealed that changes in vocal tract anatomy produce narrower formant bandwidths and spectral zeros that alter the spectral properties of EL speech. Vocal tract modeling confirmed that these spectral zeros are a function of EL placement and thus their effects will vary from user to user. Even though the addition of pitch information was associated with the greatest improvement in EL speech quality, its implementation is not currently possible because it would require access to underlying linguistic and/or neural processes. Based on these findings it was concluded that an enhancement algorithm that corrects for the low frequency deficit, the interference of the EL self-noise, the narrower formant bandwidths,