I have thought about this problem before, and I will tell you my way of thinking. Generally speaking, if you directly want to convert voice to speech, it is almost difficult to achieve, and now the neural network is not so intelligent. In other words, the speed and memory of the hardware are limited.

1. convert the voice content of the question to text;
2. realize question-answer and network realization through neural network training;
3. then convert text into speech;

I think that voice quasi-text, or text-to-speech technology, may be more complicated, if you do it yourself, it may become the core difficulty of this problem. But at the heart of the problem is step 2. Therefore, it is recommended to call third-party modules to achieve text-to-speech conversion;
go back to step 2. Do text-to-text training, at present, there are a lot of similar movie lines can be used as training data sets. Then find a more suitable neural network to complete the training quickly;

the above is what I thought before.