LOW-RESOURCE BURMESE SPEECH SYNTHESIS BASED ON VISUAL TEXT EMBEDDING AND DIFFUSION MODEL

With the advancement of deep learning and speech synthesis technology, the synthetic speech of high-resource languages such as Chinese and English has approximated natural speech closely. Burmese belongs to the Tibeto-Burman family of the Sino-Tibetan language group. Due to limitations in the scale of corpus data and depth of research, there has remained a significant gap between synthesized speech and natural speech. Traditional text embedding methods have struggled to accurately capture the complex character characteristics of Burmese, resulting in mispronunciation and unnaturalness. Under the premise of low resources, this paper proposed three methods to improve the naturalness of Burmese speech synthesis: (1) The method of "visual text embedding + phoneme text embedding" was used in the front end of speech synthesis to integrate the features of the two embedding methods to better extract complex character features and context information. (2) In order to alleviate the problem of insufficient training data due to low resources, RoBERTa pre-trained model was used in the text encoder stage. (3) In order to effectively improve the inference speed of the diffusion model, OT-CFM training and inference methods were used in the decoder. Experimental results show that by using the "visual text embedding + phoneme text embedding" method and RoBERTa pre-training model, the subjective MOS and objective MCD evaluation results of Burmese speech synthesis reach 4.17 and 7.24 respectively, which are 0.65 and 1.72 higher than the baseline model respectively. Using OT-CFM method, the inference speed is improved by 54.18% under the premise that MOS is 4.21.