ITMO and MWS AI: a speech model can learn a new language from 25 hours of recordings instead of 100
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Voice models can be extended to languages for which few recordings exist.
Scientists at ITMO and MWS AI developed SEAM, a method for adapting speech synthesis models to new languages with less training data: to reach quality comparable to the standard approach, 25 hours of audio were enough instead of 100. The method draws on knowledge the neural network already has, forming new language elements from familiar components in phonetic order. The approach was tested on Russian, Kazakh and French and on two synthesis architectures.
Sources 1