NaomiMeseret/Fine-tuning-stt-amharic-whisper

# Fine Tuning STT/TTS Models: Amharic Whisper Project This repository contains my AI Engineer task on speech processing. I focused on **speech-to-text (STT)** for **Amharic**, using the **Whisper** model family and a public **Common Voice** dataset. ## Project Summary - Task type: STT - Language: Amharic - Model family: Whisper - Main baseline: `openai/whisper-tiny` - Fine-tuned model: `openai/whisper-tiny` after a small Colab fine-tuning run - Dataset: `hadamard-2/common-voice-24-ethiopian-v2` - Main goal: compare a raw Whisper baseline with a minimally fine-tuned Whisper Tiny model on public Amharic speech ## Why I Chose This Setup I chose Whisper because it is open-source, multilingual, and widely used for speech recognition. I chose Amharic because the task encouraged under-resourced languages. I used Common Voice because it is public and provides audio-text pairs, which makes it suitable for speech-to-text experiments. I kept the notebook setup small enough to work in a Colab-style environment with limited compute. ## Repository Structure ```text Fine Tuning STT_TTS models/ ├── README.md ├── requirements.txt ├── .gitignore ├── notebooks/ │ └── amharic_whisper_common_voice_colab.ipynb │ ├── outputs/ │ ├── model_comparison.png │ └── sample_waveform.png │ └── data/ ├── raw/ │ └── README.md └── processed/ ├── README.md ├── predictions.csv └── summary.json ``` ## Deliverables This repository includes the three main deliverables requested in the task: 1. Code / Notebook `notebooks/amharic_whisper_common_voice_colab.ipynb` 2. PDF Report `reports/amharic_whisper_report.pdf` 3. GitHub Repository This project folder is organized to be uploaded as a public GitHub repository with the notebook, code, outputs, and report. ## What I Did - Loaded public Amharic speech clips from Common Voice - Ran a baseline Whisper model - Ran a minimally fine-tuned Whisper Tiny model - Compared both models on the same test samples - Saved metrics, predictions, and visua …