Enhancement of esophageal speech using statistical and neuromimetic voice conversion techniques
Résumé
This paper presents a novel approach for enhancing esophageal speech using voice conversion techniques. Esophageal speech (ES) is an alternative voice that allows a patient with no vocal cords to produce sounds after total laryngectomy: although it doesn't need any external devices, this voice sounds unnatural when compared to laryngeal speech. ES is frequently described as a harsh speech with low pitch frequency and loudness. Consequently, ES has a poor degree of intelligibility and a poor quality. To improve naturalness and intelligibility of esophageal speech, we propose a speaking-aid system enhancing ES in order to clarify and make it more natural. Given the specificity of ES, in this study, we propose to apply a new voice conversion technique taking into account the particularity of the pathological vocal apparatus. The vocal tract and excitation cepstral coefficients are separately estimated. We trained deep neural networks (DNNs) and Gaussian mixture models (GMMs) to predict "laryngeal" vocal tract features from esophageal speech. The converted cepstral vectors are then used to estimate excitation and phase coefficients by a search in the target training space previously encoded as a binary tree. The voice resynthesized sounds like a laryngeal voice, i.e., is more natural than the original ES, with an effective reconstruction of the prosodic information while retaining, and this is the highlight of our study, the characteristics of the vocal tract inherent to the source speaker. The results of voice conversion evaluated using objective and subjective experiments, validate the proposed approach.