A lot of the meaning in speech sits in the delivery, and models can now be asked about it directly.
The Hume Audio Decision API answers questions about a clip from the audio itself, with no transcript step, and returns a probability for each answer you define.
puts sighs, laughter, hesitation, overlapping speech and background sounds inside the model’s input. A transcript would have reduced all of them to the same words.
so you can ask like “Is this caller angry?”, and get a direct answer with a confidence percentage.
Most AI judges calls by reading a text transcript, which loses tone; But this one listens to the voice itself, so a sarcastic “Oh, perfect” no longer counts as happy.