MahmoudAlaa39/smart-egypt-ai-translation

# Smart Egypt AI Translation **An Integrated Artificial Intelligence Framework for Real-Time Speech-to-Speech Translation.** This repository contains the backend AI Translation Engine built for the **Smart Egypt Explorer** mobile application. The architecture deploys a highly optimized, multimodal pipeline capable of executing real-time voice translation across multiple languages natively into the colloquial Egyptian Arabic dialect (`arz`). ## Core Architecture The system utilizes Meta's **SeamlessM4T-v2-Large** model, fundamentally modified to handle real-world acoustic environments and local dialects. * **Dialect Specialist (QLoRA):** A custom Quantized Low-Rank Adapter fine-tuned on the MGB-3 corpus to map formal Arabic to Egyptian colloquialisms. * **Acoustic Denoising:** Real-time non-stationary spectral gating via `noisereduce` to suppress ambient street noise prior to matrix attention scoring. * **Semantic Memory (FAISS RAG):** An active Retrieval-Augmented Generation memory layer utilizing `all-MiniLM-L6-v2` embeddings. It intercepts ASR drift by applying deterministic zero-shot corrections with a strict L2 distance threshold (`< 0.15`). * **Zero-Disk Deployment:** Hardware I/O operations are bypassed entirely. All multi-part audio payloads execute inside volatile RAM buffers (`io.BytesIO`) using `bfloat16` precision on NVIDIA L4 instances. ## Setup & Installation **1. Clone the repository:** ```bash git clone github.com cd smart-egypt-ai-core ``` **2. Install dependencies:** ```bash pip install -r requirements.txt ``` **3. Provide Model Weights:** The base SeamlessM4T-v2-Large model weights download autonomously from Hugging Face on the first run. Place your trained QLoRA adapter in a directory named qlora_adapter_final in the project root. If this folder is missing, the API safely falls back to the base model. **4. Boot the API:** ```bash python api/main.py ``` ## Endpoints POST /translate/audio Accepts mu …