Voice models are being trained to perform "latency-budgeted reasoning," where thinking time is capped to ensure a fast response