Design and Implementation of Emotional Speech Synthesis Algorithm for Tibetan Virtual Digital People

This article first constructs an emotion classification system applicable to Tibetan, a minority language with unique phonetic characteristics and rhythmic rules, covering six emotions: joy, sadness, anger, surprise, fear, and neutrality. Using a combination of recurrent neural network (RNN) and long short-term memory (LSTM), sentiment features are extracted from Tibetan text and classified through fully connected layers. In the acoustic feature mapping stage, the fundamental frequency (F0), speech speed, volume and timbre are defined and adjusted, and emotional tags are mapped to acoustic features by using linear regression and MLP model. In order to further improve the naturalness of emotional expression, a prosody generation algorithm is designed, and the strategy of generating fundamental frequency curve and adjusting its length is optimized. In the stage of acoustic modeling and waveform synthesis, the Transformer model is used to predict the acoustic characteristics from sequence to sequence (Seq2Seq), and the WaveNet model is used to realize high-quality speech waveform synthesis. The experimental results show that the synthesized speech is excellent in naturalness and fidelity of emotional expression, especially in happiness and sadness. The research in this paper not only enriches the technical accumulation in the field of Tibetan speech synthesis, but also provides a useful reference for the study of emotional speech synthesis in other minority languages.