Researchers from ITMO University and MWS AI have introduced a new method called SEAM. This approach enables the adaptation of speech synthesis models to additional languages while significantly reducing the volume of training data required for the process.
Standard training for voice neural networks relies on text-audio pairs. The SEAM method optimizes this by utilizing the model's existing knowledge, constructing new linguistic elements from familiar components based on phonetic order. This technique addresses the challenge of limited labeled datasets for various languages.
The team validated the method using two speech synthesis architectures across Russian, French, and Kazakh. Experiments demonstrated that the approach achieves quality comparable to traditional methods using only 25 hours of audio recordings, whereas standard training typically demands 100 hours.
The research findings have been accepted for presentation at the EMNLP 2026 conference. Alexander Bukhanovsky, head of the ITMO AI Institute, noted that while this knowledge transfer reduces data requirements, it may occasionally lead to pronunciation distortions or non-standard linguistic forms.
Potential applications for this technology include the development of voice assistants, automated customer support systems, and professional voice-over services that require rapid expansion into new language markets.
