Research on end to end low resource speech synthesis based on meta learning

With the development of deep learning, much progress has been made in end-to-end speech synthesis. The success is obvious especially in quality, pronunciation accuracy, and fluency ,the speech is close to human’s, its accuracy is below human’s tolerance. All these achievements are easy to achieve in the scene of the abundant corpus, however, there are difficulties in corpus collection for 95% languages in today's world and it is difficult to get good results when facing corpus shortage yet. We propose a meta-learning approach to solve the difficulties in low-resource scenarios by learning knowledge from other languages (with a large dataset of high quality). Our model can share information between languages,makes the model trained on a small corpus a good model using the knowledge from other languages. The speech synthesized with our model sounds more natural and achieve a higher level on quality. The resultant speech is close to human’s speech in terms of fluency. We add a parameter generator to the basic end-to-end speech system and our model has smaller parameters than the multilingual model with separate parameters for each language. Experiments show that our model can share information efficiently between languages.