Advances in low-latency streaming inference have compressed AI phone conversation response times from an initial 2-3 seconds down to under 500 milliseconds over the past two years