The voice sounds a little bit distorted, and there is often a noise in the background (especially noticeable when this noise disappears when the voice pauses). I wonder, is it model limitations or is it the problem with quality of training data?