JONAHKYAGABA/-Whisper-ASR-Fine-Tuning-for-Amharic

# 🗣️ Whisper ASR Fine-Tuning for Amharic # 🗣️ Whisper ASR Fine-Tuning for Amharic This project implements a full pipeline to fine-tune OpenAI's Whisper model on **Amharic speech data** using the Hugging Face `transformers` ecosystem. It includes data preprocessing, vocabulary generation, training configuration, evaluation metrics (WER/CER), and model pushing to the Hub with experiment tracking via Weights & Biases (W&B). --- ## 📌 Key Features - Preprocesses and filters Amharic speech datasets (`KYAGABA_AMHARIC_ALFA_DATASET_100HRS`) - Trims and splits data based on hours for train/dev - Defines a custom vocabulary for Amharic characters - Sets up Whisper tokenizer, processor, and feature extractor - Fine-tunes `openai/whisper-small` for ASR transcription - Evaluates performance using **WER** and **CER** - Tracks experiments with W&B - Pushes model, tokenizer, and processor to Hugging Face Hub --- ## 🗂️ Dataset - **Train Set**: `KYAGABA/KYAGABA_AMHARIC_ALFA_DATASET_100HRS` - **Dev Set**: `KYAGABA/amharic_cleaned_testset_verified` - Filtered by: - Duration (1–30s) - Length ratio (CTC-friendly) - Token/character cleaning and normalization --- ## 🤖 Model - Fine-tuned model: `openai/whisper-small` - Output directory pattern: