Microsoft Foundry's realtime voice API (which itself is wrapping AI models from the major players) has response times in the milliseconds.