A Speech Synthesis Method based on End-to-End Model for Tibetan multi-speaker

Abstract

"Tibetan TTS generally focuses on a single speaker or a single dialect, and there is a lack of research on Tibetan speech synthesis technology with multi-speaker. This paper proposes a speech synthesis method, which can synthesize the Tibetan speech for different speakers. First, the Wiley transliteration program is used to convert Tibetan text into Latin letters. This step will effectively enhance the effect of model learning and reduce the required corpus data set and the workload of text data processing. The following step is to establish a sequence-to-sequence feature prediction network model, extract the Tibetan Latin letters’ feature vector through text preprocessing, map the vector to the mel spectrogram, and combine different WaveNet vocoders trained by different speakers to combine mel spectrogram synthesizes time-domain waveforms. The subjective and objective experimental results show that the speech synthesis effect obtained by