Theoretically - if the change was deeper than just restricting output - you could feed it thousands worth of audio tokens and get a structured response in milliseconds rather than minutes?

Say classify recorded interviews into viable/not viable candidate?