AshrafMMahdy/futo-egyptian-arabic-models

Egyptian Arabic voice input models for FUTO Keyboard

Offline Egyptian Arabic (+ English code-switching) Whisper models, adapted with FUTO's ACFT method and converted to whisper.cpp GGML, ready to import into FUTO Keyboard / FUTO Voice Input on Android.

Without ACFT, Whisper models loop or repeat on clips shorter than ~15 s inside FUTO. These don't.

Files

File	Base	Params	Size	Notes
small_egy_acft_q8_0.bin	IbrahimAmin/code-switched-egyptian-arabic-whisper-small	244M	253 MB	Runs on any modern phone. Recommended default.
small_egy_acft_q5_0.bin	same	244M	168 MB	Smaller, slightly less accurate.
turbo_egy_acft_q5_0.bin	mohammedaly22/whisper-large-v3-turbo-egyptian-code-switching	809M	~570 MB	Noticeably better on dialect and س/ص style ambiguity. Flagship phones (Snapdragon 8 Gen 3 / 8 Elite, Dimensity 9300+) with 12 GB+ RAM.
turbo_egy_acft_q8_0.bin	same	809M	~870 MB	Highest quality, slowest.

How to use on Android

1. Install FUTO Keyboard and add Arabic under Settings > Languages.
2. Download one .bin to the phone.
3. Open it with FUTO Keyboard (file manager > Open with), or Settings > Languages > Import from file. Assign to Arabic.
4. Tap the mic and speak.

How they were made

ACFT (audio context fine-tuning, FUTO's method): 2000 steps, lr 1e-6, MSE between decoder hidden states of the student (dynamic audio context) and a frozen reference (full 30 s context), on ~4 h of Egyptian speech from MAdel121/arabic-egy-cleaned. For turbo only the encoder was trained (decoder frozen, 8-bit Adam) to fit a single T4. Then whisper.cpp convert-h5-to-ggml.py and whisper-quantize.

Credit to the fine-tune authors (IbrahimAmin, mohammedaly22) and to FUTO for ACFT. Built in one Colab session; issues and better Egyptian fine-tunes welcome.