Regeneration of Speech in Voice-Loss Patients

Abstract

This paper considers regeneration of natural sounding speech from whisper-speech, produced by patients with vocal tract lesions affecting the glottis. Such reconstruction is important for both total and partial laryngectomy patients to improve on the monotonous robotized sound typical of electrolarynx devices. Reconstruction of speech from whispers has been demonstrated previously, however the resulting speech does not exhibit particularly high intelligibility, and more importantly, sounds un-natural. It is the conjecture of the authors that limited pitch variations in the reconstructed speech contributes most to that lack of naturalness. In this paper, a method for pitch contour variation in reconstructed speech is presented. This method extracts voice factors which are important to 'naturalness' from the whispered signal and applies these to the reconstructed speech. The method is based upon our previous published work which implemented an analysis-by-synthesis approach to voice reconstruction using a modified CELP codec.