A preliminary study on applying character-level RNNLM to text-to-speech system

High quality linguistic features extraction is the key component to the success of speech synthesis. However, traditional linguistic feature extraction methods are all based on natural language processing (NLP) frontends that relies heavily on feature engineering and ignores backend synthesis errors in feedback. In order to approach the goal of establishing an end-to-end speech synthesis system, in this paper, a character-level recurrent neural networks (RNNs) is adopted to directly convert input character sequences into latent linguistic feature vectors. Experimental results showed that proposed approach provides comparable performance with conventional NLP parser-based methods.