derak-isaack/Swahili-ASR-Zindi

Finetuning LLM on Zindi swahili dataset # Swahili Automatic Speech Recognition (ASR) - Zindi Hackathon Project --- ## Project Description This repository contains a Google Colab notebook developed for a Zindi hackathon focused on fine-tuning language models for Swahili Automatic Speech Recognition (ASR) using the Hugging Face Transformers framework. The notebook demonstrates dataset loading, preprocessing, model configuration, and transcription inference for Swahili ASR tasks. It is optimized for execution in Google Colab with optional integration for Google Drive to manage audio data and model checkpoints. --- ## Project Objectives - Fine-tune transformer-based speech models for Swahili ASR - Preprocess Swahili audio datasets for training and inference - Evaluate recognition outputs for accuracy and consistency - Demonstrate low-resource language model adaptation using public tools --- ## Environment This project was executed in **Google Colab**. To ensure proper execution: - Enable **GPU runtime** - Install all required dependencies before running the notebook(First notebook cell) - (Optional) Mount Google Drive for dataset persistence --- ## Running the Notebook 1. Open the `.ipynb` file in Google Colab. 2. Install dependencies if required. 3. Mount Google Drive (optional but recommended). 4. Execute cells sequentially to replicate the ASR workflow. --- ## Notes - This work was developed as part of a **Zindi hackathon challenge** focused on **Swahili language model fine-tuning**. - The goal is to improve speech recognition performance for low-resource African languages using open-source tools. --- ## Acknowledgements - Zindi Africa for hosting the hackathon - Hugging Face for providing open-source transformer models - Community contributors working on African language AI