The problem is it's not very good and there is a vast uncanny valley. Phoneme / voice conversion is an incredibly difficult problem and not likely to be solved

The problem is it's not very good and there is a vast uncanny valley. Phoneme / voice conversion is an incredibly difficult problem and not likely to be solved with any techniques known today, least ways not cheaply. Transformers show some ability to get it right , but if you wanted to use e.g. a Seq2Seq model you'd be hard pressed to come up with enough data that would generalize.