I trust these guys know what they’re doing, but determining a whisper from a spoken word doesn’t sound like a difficult problem. I think they should have just used a logistic regression first given that it seems they’re not sure how the model works and whether the last 50 frames matter due to systemic biases in the data collection or actual trends. The graph of their posterior confuses me a bit too. Is the model predicting individually for each frame? Or is it contextualized for all previous frames up to that point? If it’s the former, their approach to picking a conclusion seems off, if it’s the latter, I would question their model.

Edit: although I suppose the graph shown could be one of the better looking cases of certainty and not reflective of their typical result