Drone Transcripterv3

Egyptian Arabic, written the way teachers speak.

Whisper fine-tuned on Egyptian classroom lectures: 10.6% word error on teachers it never heard, with English terms and numbers kept as English and digits.

0:00

هنا قال لي بص يا سيدي. Point intersection. الـ points اللي الـ curve هيقابل فيها الـ X axis. Point intersection curve and X axis. بص هنا عايز أجيب الـ zeros. قابلت فين؟ negative 2, 2. يبقى عندي هنا الـ zeros.

English termNumberA real 20-second math lecture from a teacher the model never heard in training.

Skill map

Every skill on one 0–100 scale. Accuracy is 100 minus the error rate; speed is log scale on the same laptop GPU.

Drone Transcripter v3Whisper large-v3

What we’re improving next

SpeedNow: 20× real time on a laptop GPU
A second, small model for Egyptian Arabic only, trained on v3’s own transcripts, fast enough for phones and CPUs.
Standard ArabicNow: 12.2% error (Jais-ASR: 7.5%)
Add read-aloud and broadcast Standard Arabic to the big model’s training.
Everyday EgyptianNow: 16.5% error
Add Egyptian speech from outside the classroom: talk, interviews, everyday conversation.

Same clip, three models

Red words don’t match the human transcript.

Whisper large-v357% word error

هنا قال لي بص يا سيدي بوينت انترسكشن البوينت اللي الكيرف هيقابل فيها الاكس اكسز بوينت انترسكشن كيرف اند اكس اكسز بص هنا عايز اجيب الزيروز قابلتي فين نيجاتيف تو تو يبقى عندي هنا الزيروز

NVIDIA FastConformer72% word error

هنا قالي بوس ياسيدي باينت انتش البينات اللي الكيرف هيقابل فيها الاكس اكسس باينة انتشر كيرف عند اكس اكسس بص هنا عايز اجيب الزيروز قابلت فين نيجي في توتو يبقى عندنا الزيروز

Drone Transcripter v30% word error

Try it

Up to 60 seconds. Audio is transcribed and discarded, never stored.

Benchmarks

Same clips for every model, spelling-normalized. Green marks the best score in each column.

Model	Size	LecturesWER % ↓	Everyday EgyptianMR-WER % ↓	Standard ArabicWER % ↓	English termsrecall % ↑	Numbersrecall % ↑
Drone Transcripter v3	809M	10.6	16.5	12.2	90.5	91.8
Jais-ASR	5718M	20.5	13.9	7.5	57.7	81.6
Whisper large-v3	1550M	33.5	24.2	9.5	32.1	74.5
Whisper dialectal-v2	809M	34.6	20.7	12.5	2.3	61.4
Whisper large-v3 turbo	809M	42.6	29.3	10.3	12.4	65.6
NVIDIA FastConformer	115M	37.1	23.5	8.2	0.0	59.6
Moonshine base	61M	57.9	40.4	15.0	0.7	45.3

Lectures: 600 clips from 4 teachers never seen in training.
Everyday Egyptian: MGB-3 Egyptian YouTube, 773 clips, official multi-reference WER against 4 human transcripts.
Standard Arabic: FLEURS Egyptian: Wikipedia sentences read aloud, 428 clips.
English terms: Share of the 1,135 English words in the lecture clips written in English.
Numbers: Share of the 1,108 numbers in the lecture clips transcribed correctly.

Speed

Minutes of audio per minute of compute, log scale. Hardware differs, so compare within one machine.

NVIDIA FastConformer797×
A100, batch 32
Drone Transcripter v320×
laptop RTX 3060, whisper.cpp q5_0
Whisper large-v38.6×
laptop RTX 3060, whisper.cpp q5_0
Drone Transcripter v20.7×
8 CPU threads, whisper.cpp q5_0

Run it yourself

Weights (whisper.cpp ggml, q5_0, 574 MB)Release coming soon

Convert to 16 kHz mono, then transcribe with whisper.cpp:

$ffmpeg -i lecture.mp4 -ar 16000 -ac 1 lecture.wav$whisper-cli -m ggml-drone-transcripter-v3-q5_0.bin -l ar -f lecture.wav

About 20× real time on a laptop RTX 3060. Add -osrt for subtitles.

Limits

Tuned for Egyptian classroom speech. Jais-ASR is ahead on read-aloud Standard Arabic (7.5% vs 12.2%) and everyday Egyptian (13.9% vs 16.5%).
Works in 30-second windows, like all Whisper models. Long silence or music can produce repeated phrases.
Spelling follows CODA for Egyptian Arabic, so some words differ from Standard Arabic spelling.