i was sure the prosody tagging was slowing down my speech to text, so i finally benched it instead of trusting the feeling. controlled test, medians of five runs, the two arms interleaved so the machine load could not fake a winner. every clip length, from a one and a half second note to a sixty second ramble, the delta between plain transcription and transcription with the tagger was a handful of milliseconds, inside the noise. the longest real note came out thirteen milliseconds faster with prosody on than off.
the reason is architectural. the tagger runs in parallel with the speech model and finishes five to twenty five times earlier, so it hides entirely under the speech leg. it costs about three tenths of a percent of realtime, while the speech model costs one and a half to eight percent. on the short clips the tagger looks big in ratio, twenty percent of a four second note, but it still finishes before the speech does, so it never adds a millisecond to what i feel.
the actual bottleneck was a different feature that i had already disabled the week before. the emotion model was adding hundreds of milliseconds to every note, and nobody noticed because it was part of the same pipeline. once it was off, the numbers went back to noise. the lesson is that the feel argument loses to a stopwatch, and the real cost was hiding in a feature i was not even thinking about.