Chesia-Anyika/Kiswahili-ASR-WebApp-with-Whisper

An end-to-end Automatic Speech Recognition system for Kiswahili using transfer learning on OpenAI's Whisper model. Includes preprocessing, fine-tuning on Mozilla Common Voice, and deployment via a web app for real-time transcription of audio and video files. "# Kiswahili-ASR-WebApp-with-Whisper" --- ## Currently experiencing Issues pushing the related files to github due to a cloud syncing issue. Refer to the linked gamma presentation for details on the project in the mean time gamma.app --- This repository contains the code, model configurations, and training pipeline for fine-tuning OpenAI's Whisper model for **Kiswahili Automatic Speech Recognition (ASR)**. The goal is to improve transcription accuracy for Kiswahili, a widely spoken but underrepresented language in modern ASR systems. ## 📌 Project Overview Kiswahili, spoken by over 100 million people across East Africa, suffers from low representation in popular ASR models. To address this, we fine-tuned the Whisper model, a state-of-the-art speech-to-text model, using a curated Kiswahili speech corpus. This fine-tuned model is optimized for better performance on regional accents, vocabulary, and phonetics. ### ✅ Key Features * Fine-tuning using OpenAI Whisper * Support for Swahili speech data (custom dataset) * Evaluation metrics for WER (Word Error Rate) * Scripts for training, inference, and data preprocessing * Notebook integration for ease of experimentation ## 🛠️ Project Structure ``` ├── data/ # Dataset storage and preprocessing scripts ├── models/ # Checkpoints and fine-tuned model storage ├── notebooks/ # Jupyter/Colab notebooks for exploration ├── training/ # Training and evaluation scripts ├── utils/ # Helper functions and utilities ├── requirements.txt # Python dependencies └── README.md # Project documentation ``` ## 📁 Dataset We used a curated Kiswahili speech dataset consisting of transcribed audio recordings from publicly available sources. Audio files were preprocessed (converted to `.wav`, downsampled) and split into train, validation, and test s …