High "Extra inference time" value with realtime GUI leads to quieted/cutoff words

I notice that if I set the extra inference time to a high value (most prominent at 5, but starts to appear above ~1.5) that the first words I speak will be extra quiet, if not cut out entirely. This only happens if there's a long pause between speaking (ex. waiting 7s between spoken words consistently reproduces it.) At a value of 1.5 or lower, this issue does not happen and there's no issues hearing the initial words i speak.

Is there any insight into what could be causing this? It happens across 32/40/48k models and various build parameters. @yxlllc sorry for the ping but you seem to engage with the realtime usage a lot, and i'm curious if you can reproduce this behavior.

Edit: For added context, I trim all silences out of my build audio, and have confirmed that others can reproduce this problem. Interestingly, it's the least pronounced on 32k models.

Start by reproducing the realtime GUI behavior with a high Extra inference time and long pauses between words, comparing the 32k, 40k, and 48k models. Trace the realtime audio buffering and inference path involved in that setting; done means initial words remain audible after long pauses and the behavior is covered by a regression check.