adegokeisrael/SWAHILI_ASR

# Fine-Tuning Whisper (Medium) for Swahili ASR 🇹🇿🎙️ **Automatic Speech Recognition (ASR) adaptation using Mozilla Common Voice v17** > **Project Type:** Small Machine Learning Experiment (Fine-tuning + Evaluation) > **Model:** `openai/whisper-medium` (764M parameters) > **Dataset:** Mozilla Common Voice 17.0 (Swahili) > **Core Metrics:** Word Error Rate (WER), Real-Time Factor (RTF), Peak Memory --- ## 📌 Project Overview This project presents an end-to-end machine learning experiment for fine-tuning OpenAI’s **Whisper-medium** model on the **Swahili** language using the **Mozilla Common Voice v17** dataset. The goal is to improve Whisper’s transcription quality for Swahili speech by adapting the pretrained multilingual ASR model using supervised fine-tuning. The pipeline covers: - Dataset loading and preprocessing - Audio resampling and tokenization - Fine-tuning Whisper using HuggingFace Trainer - Evaluation using WER - Inference and demonstration on test samples - Reporting of latency and compute resource metrics (RTF and peak memory) This repo is designed to be **reproducible**, **clean**, and **reviewer-friendly**. --- ## 🎯 Why This Project Matters Swahili is one of Africa’s most spoken languages, yet it remains underrepresented in many speech recognition systems. By fine-tuning Whisper on Swahili speech data, this work demonstrates how modern foundation models can be adapted to support low-resource languages, improving: - accessibility tools (voice assistants, captioning) - local language technology development - speech-driven education systems - inclusive AI systems for underrepresented communities --- ## 🧠 Model Description Whisper is a multilingual encoder-decoder transformer model trained for speech recognition and translation. - **Model used:** `openai/whisper-medium` - **Parameters:** ~764 million - **Architecture:** Transformer encoder-decoder - **Framework:** HuggingFace Transformers --- ## 📂 Dataset We use the **Swahili subset* …