Voice conversion for reconstruction of dysphonic speech

Abstract

The medical term "dysphonia" is used for voice disorders caused by functional impairment of voice organs which results in inadequate oscillatory movement of the vocal folds. Dysphonic speech is typically hoarse, rough and breathy with very weak phonation or with no phonation in severe examples (for instance, after total laryngectomy). Although there are speech therapy techniques as well as surgical and post-surgical rehabilitation methods practiced in medicine, a complete restoration of dysphonic speech is usually not possible with none of these elements. This study addresses this specific problem and aims to computationally repair the voice source and vocal tract anomalies of dysphonic speech. The system proposed hereby builds upon the Source-Filter Speech Model and approaches the task twofoldedly: (i) Voice source repair involves substituting the missing glottal excitation and feeding it to the vocal tract filter according to the linear prediction equation. (ii) Vocal tract repair algorithm makes use of Voice Conversion principles and focuses on replacing the dysphonic speech parameters with those of normal speech. The performance evaluations have shown the developed reconstruction system to succeed in imitating some characteristics of normal speech, yet remarkable improvements cannot be achieved perceptually in terms of naturalness.