همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Reducing Speaker Leakage in Voice Conversion Using Speaker-Invariant Representation

نویسندگان :

Alireza Sahragard Beirami¹ Mohammad Mehdi Homayounpour²

1- Computer Engineering Department, Amirkabir University of Technology 2- Computer Engineering Department, Amirkabir University of Technology

کلمات کلیدی :

voice conversion،speech synthesis،speaker-invariant representation،speaker leakage

چکیده :

Voice conversion is one of the most important fields in speech synthesis tasks, which aims to convert the timbre of a source speech to a reference speech. One of the most important challenges in performing voice conversion is the lack of parallel data. To handle this issue, most of the models tend to use two different encoders. One for encoding the speech content and one for encoding the speaker identity from the reference speech. However, these models suffer from the speaker leakage problem where the generated speech still contains the timbre of the source speaker. To handle this problem, we used speaker-invariant representations. To prove this, we trained TriAAN-VC, a Voice Conversion system with WavLM and WavLM-SPIN features. To assess speaker leakage, we used average Cosine similarity between the source and converted samples speaker embeddings. Objective evaluations show that using this method helps to reduce the timbre of the source speaker information in the synthesized speech.