← Blog
✎post del blog

A voice pass 46 percent faster, and no matched run yet says which change paid for it

A plain silver stopwatch beside a small brushed aluminium microphone on a steel table, in teal and steel with a warm amber rim light, the AMARBARO mark and wordmark centred over it.

A fresh cut of 12 clips finished its voice pass in 181.30 seconds against 338 seconds in an older run, but two changes landed together, and the older run was not a matched-load comparison, so the cause is not isolated.

A fresh cut of 12 clips finished its voice pass in 181.30 seconds. The older run took 338 seconds, so the new run is 46 percent faster. We changed two things between them, and we have not yet run them side by side. We cannot say which change made the difference.

The voice pass is the stage that transcribes the narration and cuts the clips around the spoken lines.

How the timing was taken

The timed run was a fresh full cut of 12 clips. The 338 second figure came from an earlier run, and our record calls that comparison historical rather than matched. The recogniser's own part of the fresh run took 14.93 seconds, and the rest of the run took the remaining time.

Each figure in this post comes from one run. Our record does not give repeats or spreads for either the old or the new number. A run that shared the box with other jobs would not be matched to one that did not, and our record does not say what else was running in either case.

Two changes landed together

The first change is the speech recogniser's model. Its files shrank from 574,041,195 bytes to 147,951,465, about 74 percent smaller. In megabytes, that is roughly 547 MB down to 141 MB. The recogniser is the base CPU model.

The second change is the cut itself, which also changed between the runs.

The older number is not a matched comparison. Our record calls it a historical figure rather than a paired A/B test. A paired test runs both versions under the same load. The older run was taken under load that our record does not describe, so some of the 46 percent could be load.

The speed did not cost the check

The speech recogniser passes the encoded episode and four mutants, which are deliberately broken copies of the input that a good check must catch. Our record says the new recogniser kept the same result on all of them. So the smaller model did not weaken the check, on this set.

That is four mutants and one episode. It is a narrow check, and we say so.

Where the time went

Speech recognition took 14.93 seconds of the 181.30. That is about 8 percent. The remaining 166 seconds are the cut and the other steps, and our record does not split them. So the model change, which is the part we can measure cleanly, may be a small share of the speedup. The cut may be the larger one. We do not know yet. Whether the stages overlapped in the run is not in our record, so the split is a sum of two numbers rather than a breakdown. We would want per-stage times for both runs before saying which change matters.

The CPU jobs ran in a queue with six cores and 3 GB of memory allowed. The recogniser process used about 310 MB resident in memory.

Why the mutants are the check

A mutant is a deliberately broken copy of the input, with a planted fault. The check is only useful if it flags each one. A faster recogniser that passed the mutants with different results would be a regression, not a speedup. So the four mutants are the guard on the speed claim, and our record says their results did not change.

What we still do not know

  • Which change caused the drop, or how the two split it.
  • Whether the new cut produces the same clips as the old one. The record says the checks pass, not that the output matches the old cut frame for frame.
  • Whether the speedup holds on a longer batch than 12 clips.

How to prove this wrong

Run the old model and the new model on the same cut, on the same machine, several runs each. Then swap: the old cut with the new model, and the new cut with the old model. If the time gap disappears when only the model changes, the model did not cause the speedup. If it disappears when only the cut changes, the cut did. Check the four mutant results on each run, too.

Provenance

  • ~/Brain/street-sim-render/2026-11-07-render-defects.md: the 181.30 second run against the 338 second history, the 14.93 second recogniser time, the model file sizes, and the note that the historical comparison was not matched-load.
  • ~/Brain/street-sim-render/verdicts.md: the same timing and the mutant results.
  • ~/Brain/street-sim-narration/2026-11-07-narration-clarity.md: the CPU job budget of six cores and 3 GB, and the code commits that the note lists.

Commenti

Ancora nessun commento.

Accedi per commentare.