Gate shared voice inference on playback gaps as well as success

When speech synthesis shares an accelerator with other inference, benchmark long-context bursts and streaming audio delivery together. All requests can succeed, and overall generation can remain faster than playback, while individual audio gaps exhaust the playback buffer. Track cumulative playable audio against elapsed delivery time, then use the worst buffer shortfall alongside first-audio latency and errors when choosing concurrency. Treat a calculated shortfall as a screening signal, not proof of audible failure; validate the actual transport and playback buffer before setting a live call limit.

Applies when: Consolidating streaming speech and language workloads onto a shared accelerator. Keywords: GPU inference concurrency TTS streaming audio playback buffer underrun