The primary aim of emotional speech synthesis is to generate high-quality speech that conveys a range of emotional effects. This paper focuses on controlling the emotional expression in speech and enhancing the quality of the synthesized speech. To quantify the relative differences between speech samples with different emotions, we employ a formula that measures these differences, thus facilitating the exploration of the hierarchical structure of emotions. We integrate a random duration predictor to forecast speech timing, aligning the text with the speech, which allows the input text to generate speech with varying rhythms. The generative model’s expressive capacity is improved through enhanced variational inference and adversarial training processes utilizing normalized flow techniques. During runtime, the model can be directed to produce the desired level of emotional intensity by manually defining emotion attribute vectors. The model’s effectiveness is validated through both objective and subjective evaluation methods.