Yes, you're probably right - I hadn't thought that the voice + noise is going to be one thing, extracting just the noise will be difficult, unique and maybe not

Yes, you're probably right - I hadn't thought that the voice + noise is going to be one thing, extracting just the noise will be difficult, unique and maybe not trainable at all. I have thought about the constructing a new reality thing - I wouldn't be surprised if models ended up being trained to misspeak words which could get confusing...