This would require a machine learning model and would be difficult to do real time voice generation. For my master’s defense i used Descript to synthesize my voice. Fed the model training data consisting of my voice and used it to generate voice for my presentation. The generation was not realtime and takes like 30 seconds for a minute of audio. It sounds like a robot version of my voice but did the trick. Maybe if you throw a lot of cloud compute at the problem the voice generation could be realtime (but at a cost of course)