Exemplar-Based Methods for Mandarin Electrolaryngeal Speech Voice Conversion

Hsin-Te Hwang, Chia-Hua Wu, Ming-Chi Yen, Yu Tsao, Hsin-Min Wang

Abstract: Electrolarynx (EL) can restore speech to patients without vocal cords, but the resulting speech often lacks naturalness and intelligibility. Inspired by recent advances in exemplar-based methods for speaker voice conversion (VC), this study explores their application to electrolaryngeal speech voice conversion (ELVC). As a baseline, we apply the k-nearest neighbors VC (KNN-VC) method with self-supervised learning (SSL) representations to ELVC. To circumvent the SSL feature mismatch between EL speech and natural speech when applying KNN-VC to ELVC, we propose two meth-ods under the parallel VC setting: the parallel KNN-ELVC method and the parallel locally linear embedding (LLE)-based ELVC method. Experimental results demonstrate that while directly applying KNN-VC to ELVC improves the naturalness of EL speech, our proposed methods, especially LLE-ELVC, significantly improve both naturalness and in-telligibility. Furthermore, LLE-ELVC is comparable to our previous deep-learning-based ELVC method, further demon-strating the potential of exemplar-based methods in ELVC tasks.