IbrahimDayax/somali-whisper-small

Evaluate and fine-tune OpenAI’s Whisper (small) for Somali automatic speech recognition (ASR) using LoRA parameter-efficient fine-tuning. This repository/notebook contains baseline evaluation, EDA on Somali ASR/TTS datasets, LoRA fine-tuning, and evaluation of the fine-tuned model. # SomaliWhisper **Project:** SomaliWhisper **Purpose:** Evaluate and fine-tune OpenAI's Whisper (small) for Somali automatic speech recognition (ASR) using LoRA parameter-efficient fine-tuning. This repository contains a Jupyter notebook with baseline evaluation, EDA on Somali ASR/TTS datasets, LoRA fine-tuning, and evaluation of the fine-tuned model. ## Table of Contents - SomaliWhisper - Table of Contents - Summary / Highlights - Repository Contents - Environment \& Installation - Datasets Used - Exploratory Data Analysis (EDA) — Key Stats - `somali_tts` (local TTS manifest) - `soomali_asr` (local ASR manifest) - Baseline (Untuned) Evaluation — Results - Per-dataset results (baseline) - LoRA Fine-tuning — Training Summary - LoRA Evaluation — Results - Per-dataset results (LoRA-finetuned) - Comparison \& Interpretation - Caveats, Warnings \& Known Issues - Reproducibility / How to Run (Quick Start) - Recommended Next Steps - Files / Repository Organization - Acknowledgements \& License --- ## Summary / Highlights - Performed baseline evaluation of `openai/whisper-small` on two Somali datasets from Hugging Face. - Performed EDA across local Somali TTS/ASR manifests (text and audio summaries). - Trained a LoRA adapter on top of `whisper-small` (parameter-efficient training). - After LoRA fine-tuning, the model showed **substantial improvements** on the small `adityaedy01/somali-voice` subset and measurable improvements in CER for `nurfarah57/somali_asr`. Results are based on the actual runs and logs below. - Important caution: some evaluation splits are extremely small (2–3 samples), so metrics should be treated as illustrative, not definitive. --- ## Repository Contents This repository contains: - **`SomaliWhisper.ipynb`** — Main Jupyter notebook covering the complete pipeline - **Google Drive Resources:** - **Datasets:** Training datasets and manifests - **Model Directory:** Trained LoRA adapters and outputs The notebook includes: - Dependencies installa …