Synthesis Speech Based Data Augmentation for Low Resource Children ASR

Successful speech recognition for children requires large training data with sufficient speaker variability. The collection of such a database children’s voices is challenging and very expensive zero/low resource language like Punjabi. In this paper, the scarcity issue low resourced Punjabi addressed through two levels augmentation. original corpus first augmented by modifying prosody parameters pitch speaking rate. Our results show that augmentation improves system performance over baseline system. Then combined used to train TTS generate synthesis extended dataset further generating utterances using text-to-speech sampling model methods increase acoustic lexical diversity. final indicates relative improvement 50.10% 57.40% diversity based in comparison respectively.