Termuxphone/dialects-ai-foundry-mobile

Algerian Dialect Classifier & Darja ASR (Offline Edge AI)

Quantized, offline-first speech pipeline designed to run entirely on Android CPUs (Termux) without internet connectivity.

Included Quantized Models (273 MB Total)

1. 8-Wilaya Dialect Classifier (dialect_onnx_int8/): Quantized ONNX INT8 (92 MB) via onnxruntime.
2. Algerian Darja ASR (ggml-whisper-dz-q5_1.bin): Fine-tuned Whisper quantized to GGML Q5_1 (181 MB) via whisper.cpp.

Technical Benchmark: dialects-ai-foundry vs. Cloud APIs

Feature / Metric	Cloud Speech APIs (OpenAI / Google)	dialects-ai-foundry (Mobile / Edge)
Internet Connectivity	Mandatory (Fails offline)	100% Offline (Air-Gapped / Airplane Mode)
Data Privacy	Audio sent to foreign servers	Zero-Cloud Privacy (Audio stays on device)
Wilaya Dialect Routing	Not supported	Built-in 8-Wilaya Classifier (ONNX INT8 - 92 MB)
Algerian Darja Accuracy	Forces Modern Standard Arabic	Fine-tuned on Algerian Darja (GGML Q5_1 - 181 MB)
Operational Cost	Recurring USD billing per minute	Zero runtime cost (Runs on standard CPU)
Footprint	Heavy data-center GPUs	273 MB total on Android CPU

Quick Start

Clone and run from GitHub: https://github.com/attitawfikatti-alt/dialects-ai-foundry

Commercial Licensing & Enterprise Solutions

Released under CC BY-NC 4.0 for non-commercial research. For commercial licenses, on-premise deployments, or custom fine-tuning:

• Email: attitawfikatti@gmail.com
• GitHub: https://github.com/attitawfikatti-alt/dialects-ai-foundry