That's a really intriguing idea. I wonder if you could go a step further and include a teensy amount of model data tailored to the original speaker(s). i.e. Te

That's a really intriguing idea. I wonder if you could go a step further and include a teensy amount of model data tailored to the original speaker(s). i.e. Telling the reconstruction algorithm "here's how he speaks" and "here's what he's saying".