It’s the former. It’s looking at audio data

So you can ask it sentiment, determine if someone is giggling, crying, laughing, can maybe even detect a condescending tone or flirtatious tone etc.