“Russ d'Sa predicts that future voice AI models will process audio directly, but progress is currently limited by a lack of sufficient audio-based conversational training data.”