汉语儿童情感语声合成

Affective speech synthesis of Chinese children

情感语声合成技术对于人机交互具有重要的意义。面对儿童情感语声合成所需汉语语声数据资源缺乏以及模型训练时长较长等问题,该文提出利用迁移学习实现汉语儿童情感语声合成的方法。首先基于汉语语声数据库训练深度学习模型实现中文语声端到端合成模型,再使用高质量大样本的中文情感语料库完成情感语声合成模型,最后利用自行采样的小样本汉语儿童情感语料对模型进行迁移学习实现低资源的语声合成。客观实验结果中梅尔倒谱失真指标为4.91,主观听辨实验指标分别为3.61和4.17。通过实验对比表明,该文的方法在情感语声合成技术的应用上具有良好的性能表现,并且优于现有先进的低资源情感语声合成方法。

Emotional speech synthesis technology is of great significance for human-computer interaction.Facing the lack of Chinese speech data resources required for children’s emotional speech synthesis and the long time of model training,this paper proposes a method of using transfer learning to realize Chinese children’s emotional speech synthesis.This paper first implements the Chinese speech end-to-end synthesis model based on the Chinese speech database training depth learning model,then uses the high-quality and large sample Chinese emotional corpus to complete the emotional speech synthesis model,and finally uses the self sampled small sample Chinese children’s emotional corpus to transfer the model to realize low resource speech synthesis.The objective experimental results show that the Mel cepstrum distortion index is 4.91,and the subjective auditory discrimination experimental indexes are 3.61 and 4.17 respectively.The experimental comparison shows that the method in this paper has good performance in the application of emotional speech synthesis technology,and is better than the existing advanced low resource emotional speech synthesis methods.