I think many heavy TTS users (including myself) slowly train to use higher speeds after which point nothing sounds particularly natural. What I want is trained speech models that remain coherent at high speeds (over 3x). Even better if there's bi/multi lingual models that can seemelessly switch between languages.