Comparative Study for Multi-Speaker Mongolian TTS with a New Corpus

Low-resource text-to-speech synthesis is a very promising research direction. Mongolian the official language of Inner Mongolia Autonomous Region and spoken by more than 10 million people worldwide. Mongolian, as representative low-resource language, has relative lack open-source datasets for its TTS. Therefore, we make public an multi-speaker TTS dataset, named MnTTS2, related researchers. In this work, invited three announcers to record topic-rich speeches. Each announcer recorded h speech, whole dataset was 30 in total. addition, built two baseline systems based on state-of-the-art neural architectures, including Fastspeech 2 model with HiFi-GAN vocoder full end-to-end VITS multi-speakers. On system FastSpeech2+HiFi-GAN, speakers scored 4.0 or higher both naturalness evaluation speaker similarity. achieved scores 4.5 similarity scores. The experimental results show that published MnTTS2 can be used build robust models.