Electrolaryngeal (EL) speech enables laryngectomees to resume oral communication, but its mechanical noise and nearly fixed pitch often reduce intelligibility and naturalness. This talk presents three studies on sequence-to-sequence EL voice conversion (ELVC). First, we replace alignment-dependent frame-wise mapping with a Transformer-based model initialized through text-to-speech pretraining. Second, we extend the framework to speech from real patients using an intermediate pretraining stage on simulated EL speech, a speech-encoder loss for preserving phonetic content, and duration-based data augmentation. Finally, in a preclinical study of a novel nasal electrolarynx, we show that the effectiveness of acoustic representations depends on the EL device and that synthetic parallel data can mitigate data scarcity.

Ming-Chi Yen is a postdoctoral fellow at the Institute of Information Science (IIS), Academia Sinica, where he worked as a research assistant from October 2012 until assuming his current position in August 2026. He received his Ph.D. in Computer Science and Information Engineering from National Taiwan University in 2026 and his B.S. and M.S. degrees from National Chiayi University in 2009 and 2012, respectively. His research focuses on speech signal processing, particularly atypical speech. His current work applies voice conversion techniques to improve the intelligibility and naturalness of electrolaryngeal speech, while his broader research interests include dysarthric speech, speech representation learning, and sequence-to-sequence modeling. His work has appeared at IEEE ASRU, Interspeech, and APSIPA ASC, as well as in IEEE Internet of Things Magazine.