Abstract: Speech synthesis is one of the key technologies of voice interaction on intelligent home appliances. The quality of synthesized speech directly influences the users’ intelligent interactive experiences. For the problem that the duration of synthesized speech is deterministic and lack of rhythms, utilizing the flow-based stochastic duration predictor to optimize the mainstream Glow TTS model, and an experiment was conducted with Japanese as the research object. The result indicates that the accuracy of speech duration prediction and the quality of synthesized speech are greatly improved. It can provide guidance for the following optimization study on speech synthesis of intelligent home appliances.

Key words: Intelligent home appliances, Speech synthesis, Glow TTS model optimization, Standard flow, Stochastic duration predictor