The STC text-to-speech system for Blizzard Challenge 2019

The paper presents text-to-speech system developed at STC for the Blizzard Challenge 2019. This year, the task is to build a TTS system for Mandarin Chinese using found data suitable for expressive TTS. Provided corpus contains 8 hours of speech by a native speaker with text annotations. We describe a neural speech synthesis system for Mandarin Chinese built without any significant prior knowledge about the language. Input text is converted to a sequence of phones using publicly available tools. Then, a sequence of phones is turned into a spectrogram by a Tacotron-based neural network. Finally, the spectrogram is converted into a waveform using a LPCNetbased neural network. Our system is based on learning deep representations and does not explicitly use or predict such features as pitch, duration of every phone, etc. We also discuss our system’s performance in listening tests conducted by organizers of the challenge.