This is an individual Natural Language Processing (NLP) project focused on developing an end-to-end speech recognition tool for the low-resource Dholuo language using OpenAIβs Whisper model.
# π¬ Dholuo Automatic Speech Recognition (ASR) Project
*Empowering Low-Resource Language Access with Whisper | NLP + Web App Development*
## π Nairobi, Kenya | Sept 2025 β Ongoing
---
## π§ Overview
This is an **individual NLP research project** focused on building an **end-to-end Automatic Speech Recognition (ASR) tool** for **Dholuo**, a low-resource language spoken in Kenya. The system is powered by **OpenAIβs Whisper model**, fine-tuned to better understand Dholuo phonetics and vocabulary.
> π£οΈ The project aims to enhance **language accessibility**, **education**, and **digital documentation** for underrepresented African languages.
---
## π― Key Objectives
- π₯ Collect & preprocess **Dholuo speech data** using Mozilla Common Voice
- π§½ Normalize and clean transcription data (tokenization, casing, punctuation)
- π§ Fine-tune **Whisper** via transfer learning for Dholuo-specific ASR
- π Evaluate model accuracy with **WER** & **CER**
- π Deploy a live **Streamlit web app** for transcription
- π§ͺ Provide a **Gradio prototype** for real-time testing
---
## π οΈ Tech Stack & Tools
| Task | Tools / Frameworks |
|--------------------------|-------------------------------------------------------------|
| Data Collection | `Mozilla Common Voice`, `Python scripts` |
| Preprocessing | `NLTK`, `NumPy`, `re`, `Whisper-compatible text cleaning` |
| Model Training | `Whisper`, `Google Colab`, `PyTorch`, `Transfer Learning` |
| Prototyping | `Gradio`, `Streamlit` |
| Evaluation | `WER`, `CER`, `scikit-learn`, `JiWER` |
---
## π Project Structure
```
dholuo-asr/
β
βββ 1. ASR Project: Filtering ASR tsv files/
βββ 2. ASR Project - File Conversion/
βββ 3. ASR Projects - Mapping & Filtering/
βββ 4. ASR Project - Preprocessing/
βββ 5. ASR Project - Training/
βββ README.md
```
---
## π β¦