That's not an objective measurement of how easy he is to understand for humans. That's a predictive score from a model trained on some set of audio recordings

That's not an objective measurement of how easy he is to understand for humans. That's a predictive score from a model trained on some set of audio recordings of how likely it would be for it to have produced that audio. This can only possibly be construed as a "measurement" with respect to the model. It's a measure of how the model relates to the input. I'd claim that's not a good proxy for human comprehension at all. In text, it works a good bit better as text data and grammar are far less rich than speech with intonation, accent, cadence, etc.