abeladamushumet/Fluentian_STT_TTS_Project

Amharic STT & TTS pipeline using Whisper-small and Coqui TTS # Fluentian STT/TTS Project **Fine-Tuning Speech-to-Text Model for Amharic Language** A comprehensive implementation of Speech-to-Text (STT) system using OpenAI Whisper, fine-tuned on Amharic speech data from the Leyu dataset. This project was developed as part of the Fluentian Internship Programme - Task Round 1. --- ## 📋 Table of Contents - Project Overview - Task Assignment - Goals - Model Selection - Dataset - Project Structure - Setup & Installation - Usage - Training Process - Results - Challenges & Solutions - Low-Compute Fine-Tuning - Deployment Considerations - Key Learnings - Author --- ## 🎯 Project Overview This project explores and experiments with Speech-to-Text (STT) systems by fine-tuning OpenAI's Whisper model on Amharic, an under-resourced language. The implementation demonstrates the complete ML pipeline from data preprocessing to model evaluation, with a focus on handling real-world challenges in low-resource language ASR systems. **Key Highlights:** - Fine-tuned Whisper-small model on 1000 Amharic speech samples - Achieved significant WER improvement: 1.4310 → 1.0300 - Handled Gojjam dialect variations for realistic ASR scenarios - Complete end-to-end pipeline with preprocessing, training, and evaluation --- ## 📝 Task Assignment **Fluentian Internship Programme - Task Round 1: AI Engineer Task** **Task Title:** Fine-Tuning STT/TTS Models **Objective:** Explore, experiment, and report on speech processing models (Speech-to-Text and/or Text-to-Speech) using open-source models and public datasets. **Requirements:** 1. Select at least one open-source STT or TTS model 2. Find a compatible public dataset (encouraged: under-resourced languages) 3. Run the model and perform minimal fine-tuning if feasible 4. Document the entire process with detailed analysis **Deliverables:** - Working code/notebook demonstrating STT/TTS - Comprehensive PDF report addressing all evaluation criteria - Public GitHub repository with code and documentation * …