I sometimes feel like its a rough transcript of the audio or similar. Because not only is the response apparently natively audio, but I've often seen the text adhere to the audio only as a best effort rather than a full on accurate representation of the audio.