Multi-speaker TTS system for low-resource language using cross-lingual transfer learning and data augmentation

We propose a multi-speaker text-to-speech (TTS) system for use with low-resource languages when only a very small amount of target language data is available. We investigate the effects of using high-resource language datasets and augmented data during model training, and compare different strategies for fine-tuning the model. After using a combination of cross-lingual transfer learning, a small amount of target language data and augmented target language data for training our low-resource TTS model, we then fine-tuned the pre-trained model using the original and augmented target language data. Our experimental results show that by sequentially training the model with high-resource language data, target language data and augmented target language data, followed by gradual fine-tuning using the original and augmented target language data, our system was able to achieve the most natural speech after text-to-speech conversion, achieving a native speaker mean opinion score of 3.50 on a scale of 0 to 5.