Statistical Voice Conversion Techniques for Alaryngeal Speech Enhancement

This position paper gives a brief overview of our developed technologies for enhancing alaryngeal speech (AL speech) uttered by laryngectomees. There are several alternative speaking methods for laryngectomees to produce AL speech. However, any type of AL speech suffers from lack of naturalness and speaker individuality (identity). To address this issue, we have developed statistical voice conversion techniques for AL speech enhancement. Our developed techniques are capable of converting AL speech into normal speech in a probabilistic manner on the basis of statistics extracted from training data consisting of utterance pairs of AL speech and normal speech. Moreover, they are also capable of flexibly controlling voice quality of enhanced speech to effectively recover speaker individuality. We have developed three AL speech enhancement systems for 1) esophageal speech, 2) electrolaryngeal speech, and 3) bodyconducted silent electrolaryngeal speech. Our experimental results have demonstrated that these systems yield significant improvements in naturalness and speaker individuality of each type of AL speech.