The problem
The school corrected pronunciation by hand. Each teacher listened to dozens of audios per class and wrote up the feedback. As volume grew, the bottleneck wasn’t pedagogical — it was time.
What was built
A three-mode system (individual, test, and batch) where the teacher submits the class’s audios and receives, per student, what was said, what should have been said, and the feedback. The language model is only called when it’s worth it: an acoustic screening step runs first, using a deliberately weak speech recognizer, and discards the audios where the student clearly got it right.
That screening step was the discovery that changed the math. A better “ear” understands the student despite the accent — and the error disappears. The sensor needs to be worse than the model for the filter to work. Confirmed five times, with different models.
What was measured
- Cost per 100 audios: R$ 2.29 → R$ 0.55 → R$ 0.30 → R$ 0.26, across three rounds of optimization (screening, batch repackaging, production confirmation).
- On a typical day: 181 feedbacks, 271 batches, 41,651 audios screened, zero errors, zero fail-opens in three days.
- Median queue wait: 7 minutes, with three machines transcribing.
- The production model was tested against the vendor’s next family in a blind-listening bench: the new family failed and production stayed where it was — a recorded decision, not an opinion.
What the client didn’t have to do
Switch platforms, install anything on the teachers’ computers, or learn a new tool. The system is a single page.