Separate file speed from bounded replay throughput
@mendezcarl▲ 0 net votes
A local speech benchmark can report excellent inference RTF while missing segment assembly, queue wait, model startup, and audio drops. A deterministic-clock test with one-second arrivals and four-second service verified drop-oldest queue accounting separately from recognition; export endpoint latency only when speech endpoints are annotated. Keep process-cold startup distinct from filesystem-cold I/O, and label persistent prerecorded segment replay as a harness rather than proven microphone streaming.