First Step Towards End-to-End Parametric TTS Synthesis: Generating Spectral Parameters with Neural Attention

In conventional neural networks (NN) based parametric text-to-speech (TTS) synthesis frameworks, text analysis and acoustic modeling are typically processed separately, leading to some limitations. On one hand, much significant human expertise is normally required in text analysis, which presents a laborious task for researchers; on the other hand, training of the NN-based acoustic models still relies on the hidden Markov model (HMM) to obtain frame-level alignments. This acquisition process normally goes through multiple complicated stages. The complex pipeline makes constructing a NN-based parametric TTS system a challenging task. This paper attempts to bypass these limitations using a novel end-to-end parametric TTS synthesis framework, i.e. the text analysis and acoustic modeling are integrated together employing an attention-based recurrent neural network. Thus the alignments can be learned automatically. Preliminary experimental results show that the proposed system can generate moderately smooth spectral parameters and synthesize fairly intelligible speech on short utterances (less than 8 Chinese characters).