Researchers from ITMO University and MWS AI, a division of MTS Web Services, have developed a new method called SEAM for adapting speech synthesis models to new languages. The approach allows neural networks to utilize existing knowledge rather than learning phonetic elements from scratch.
The method was tested on Russian, Kazakh, and French languages using two different speech synthesis architectures. Experimental results demonstrated that SEAM maintains high synthesis quality while requiring only 25 hours of audio data, compared to 100 hours typically needed by standard approaches.
This technology is designed to assist companies developing voice assistants, dubbing services, and customer support automation tools. It facilitates faster adaptation to new markets by reducing the volume of annotated speech data required for training.
Stamatis Lefkimiatis, head of fundamental research at MWS AI, was recognized in the Top 2% Scientists international database in October 2026. He ranked in the top 1% of researchers in the Artificial Intelligence and Image Processing category based on citation metrics excluding self-citations.