Drone Transcripter 4
809M-parameter Whisper large-v3-turbo, fully fine-tuned by Drone Academy for Egyptian Arabic: classroom lectures first, then everyday speech, Modern Standard Arabic and Arabic/English code-switching. Trained on 329.5 h of human-reviewed and Gemini-cleaned Egyptian audio (75,390 clips) in two epochs on one A100.
Try it live: https://asr.drone-academy.online.
Results (WER, lower is better)
| set | this model | base large-v3-turbo | notes |
|---|---|---|---|
| Egyptian lectures, held-out (600 clips) | 10.2 | 64.0 | our target domain |
| MGB-3 Egyptian, multi-reference (773 clips) | 15.3 | GigaSpeech 3 scores 17.99 on the same clips | |
| FLEURS ar_eg (MSA) | 10.5 | ||
| Casablanca (Egyptian) | 40.2 | everyday speech; GigaSpeech 3 38.1 | |
| GigaSpeechBench EGY | 48.46 | official scorer, with punctuation | |
| English words inside Arabic speech | 90.1 % correct | ||
| Numbers | 91.7 % correct |
Files
| file | use |
|---|---|
drone-transcripter-v4/ |
Hugging Face transformers checkpoint (safetensors), WhisperForConditionalGeneration |
ggml-drone-transcripter-v4-f16.bin |
whisper.cpp model (f16), runs on CPU or Metal/CUDA via whisper.cpp |
ggml-drone-transcripter-v4-q5_0.bin |
whisper.cpp model quantised to q5_0 (574 MB); same accuracy class, 2.8× smaller, the one to download for laptops |
Use with transformers
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="AbdelrhmanX2/drone-transcripter-4", device="cuda")
print(asr("clip.wav", generate_kwargs={"language": "ar", "task": "transcribe"})["text"])
Use with whisper.cpp
whisper-cli -m ggml-drone-transcripter-v4-q5_0.bin -l ar -f clip.wav
Training
Base openai/whisper-large-v3-turbo, full fine-tune, 2 epochs, 4,712 steps, batch 32, lr 1e-5 with 5 % warm-up and
linear decay, bf16, one A100-40GB (~3.5 h). Data: 122.4 h of human-reviewed lectures plus 207.1 h of platform
lectures re-cut and cleaned with Gemini. Songs and music excluded.
Limitations
Strongest on lectures and classroom speech. Casual conversation and noisy phone audio are weaker (Casablanca row). Outputs are unpunctuated Egyptian Arabic in Arabic script; English words are kept in Latin script.
Licence and citation
Apache-2.0, matching the base model. Please cite Drone Academy and link to this page.