Where Drone Transcripter is going, and where it stands today. · Updated Nov 1, 2026
Word error rate: lower is better.
Train the small model once on today’s labels to prove every step runs end to end.
Done when: Training runs, checkpoints save.
Ran to step 3,600; checkpoints saved.
Blind A/B review of 100 clips: v3’s text vs Gemini’s, without knowing which is which.
Done when: Gemini stays at about 1–2% word error.
9 of 100 reviewed.
Re-cut platform clips at pauses (no half words), clean all 219 h with Gemini, audit 2% by hand.
Done when: 219 clean hours.
The big model on reviewed + clean data, Standard Arabic and bad-mic augmentation. New held-out bad-mic test.
Done when: Beats v3 on lectures and on bad mics.
Egyptian-only FastConformer on the same clean data, fast enough for phones and CPUs.
Done when: Close to v4’s accuracy at many times the speed.
Swap v4 into the app, update this site’s scores, then release the weights.
Done when: v4 live.
| Human-reviewed lectures | 122 h | Clean |
| Platform lectures (auto-labelled) | 219 h | Needs re-cut + clean-up |
| Public Egyptian speech | 94 h | Auto-labelled by v3 |