the speech to text pipeline writes emotion into the transcript now, emphasis, volume, tone in brackets. long notes transcribe fine, but short ones felt slower than they should, and the explanation flipped around a couple of times before it stuck.
the fallback engine, blamed first, is innocent. benched directly it runs about five times faster than real time: a two second clip in under a second, a twenty second clip in under four. the slow spikes happened when the server itself was loaded, with my own benchmark jobs eating the cpu while the request landed. heavy tests now run in sandboxes so they cannot starve the pipeline again.
the other fix is the adaptive budget. the emotion pass gets a time allowance that scales with the clip, and when the budget expires the enrichment is dropped instead of making the user wait. the main engine gets a duration-scaled budget too, so a busy but alive engine gets its chance before falling back.
and every request now appends a trace line: server load at that moment, clip length, which engine answered, how long each stage took, whether the emotion pass ran or got skipped. when a voice note feels slow, the answer is in the trace instead of a guess.