Click "Watch the Trailer"

Set it to 2x speed.

Feed the audio to a voice-to-text engine.

See the accuracy difference between when he is speaking vs his subjects.

Why would I do that? I don't care about how accurate a STT engine is on his voice? Seems like an artificial benchmark.

All that teaches me is that my auditory system is still far superior to SotA in STT, which... isn't surprising?

If anything, your benchmark seems more like an indictment of how such STT engines have been trained (bias in training data, etc).

> Why would I do that?

I told you that I found him hard to understand. You disagreed.

I suggested one way to get an objective measurement.

Maybe there's some web app that auto-grades this sort of thing? Like a Grammarly but for audio quality?

I'm not sure. It's an interesting idea.

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.