Transfer Learning for the Russian Language Speech Synthesis
Aleksander D. Kuzmin, Sergey A. Ivanov · 2021
In this article, we suggest a method for creating a high-quality model of speech synthesis for the Russian language. Even though the task of speech synthesis is widespread, only large corporations can create high-quality models of speech synthesis since this requires a large amount of well-processed data recorded by the voice of one person. The proposed method allows the creation of a speech synthesis model, using about 30 minutes of target speech and about 20 hours of speech from different people from open sources. This model will be comparable in quality to the models of large companies that were trained on large amounts of data.