@kaia elevenlabs voices made form actual speech samples are usually better at capturing the speech style. for base models, you often have to crank up the 'expressiveness' slider and that can become very unnatural too, so you have to regenerate until you get a good take.