SpixerD/Amazighini

Fine-tuned Speech transcription and translation model for speech to text from amazigh as a source language. # Amazighini: Fine-Tuning SeamlessM4T for Amazigh to Arabic and English Translation ## Overview Amazighini is an AI-powered project focusing on preserving Amazigh heritage by translating Tachelhit Amazigh dialect speech to Arabic text. The main use cases for this project are educational purposes and facilitating communication between rescue teams and Amazigh speakers in rural areas during disasters. This project fine-tunes the *SeamlessM4T* model using Hugging Face's Transformers library to improve translation accuracy and efficiency. ## Table of Contents - Project Structure - Data Pipeline - Data Preparation - Model Fine-Tuning - Evaluation - Usage - Future Enhancements - Demo - Links - Contributing - License ## Project Structure The main files and directories in this project are: - `fine-tuning_seamless.ipynb`: Jupyter notebook containing the code for fine-tuning the SeamlessM4T model. - `fine-tuned-seamless-m4t/`: Directory where the fine-tuned models will be saved. - `dataset_amazigh/`: Directory containing the collected audio recordings and annotations. ## Data Pipeline The data for this project was collected from Amazigh speakers who recorded Arabic phrases translated into Tachelhit Amazigh. The phrases focused on common small talk and frequently used phrases by rescue teams. ### Steps: 1. **Data Collection**: Providing Arabic phrases to Amazigh speakers who recorded their translations in Tachelhit Amazigh. 2. **Organization**: Saving the recorded audio files in a `wav` folder and the corresponding annotations in an `annotations` folder. Each annotation file is a CSV that includes the WAV filenames and their translations in English and Arabic. 3. **Dataset Compilation**: Organizing the entire collection of audio recordings and annotations into a directory named `dataset_amazigh`. ## Data Preparation The data preparation process involves loading the CSV files into pandas DataFrames, adding full paths to the filenames, and converting the DataFr …