amina-mardiyyah/ASR-Nigeria-Pidgin

# Towards End-to-End Training of Automatic Speech Recognition for Nigerian Pidgin ## Overview This project explores **end-to-end training of Automatic Speech Recognition (ASR) systems for Nigerian Pidgin**, a widely creole spoken language across West Africa but underrepresented in speech technologies. Despite being spoken by over **75 million people**, the Pidgin language lacks sufficient digital resources and language technologies. This work addresses the gap by building datasets and training ASR models tailored for a variant of the pidgin language - Nigerian Pidgin, contributing towards more inclusive and equitable language technologies. - **Project Website:** ASR Nigerian Pidgin - **Interactive Demo:** Hugging Face Space - **Dataset:** Hugging Face Datasets - **Paper:** Arvix --- >**Towards End-to-End Training of Automatic Speech Recognition for Nigerian Pidgin**\ >Amina Mardiyyah Rufai, Abeeb Afolabi, Daniel Ajisafe, Oluwabukola Adegboro, Esther Oduntan, and Tayo Arulogun ---- ## Problem Most state-of-the-art ASR systems are trained on high-resource languages (e.g., English, German, Japanese). This creates a significant gap for low-resource languages like Nigerian Pidgin, where lack of data and research limits the development of speech-enabled applications. **Goal:** This project focuses on the development of an end-to-end speech recognition system customised for Nigerian Pidgin English. --- ## Approach We implemented and compared three(3) end-to-end ASR training methods, experimenting with different ASR pipelines and architectures: 1. **Wav2Vec2-Base** 2. **Wav2Vec2-XLSR53** 3. **NVIDIA NeMo** Each model was trained and evaluated on our Nigerian Pidgin speech corpus. Our final and best performing model on the test set was the **Wav2Vec2-XLSR53**, achieving a reduced error rate of 29.6%. ## Key Contributions - A publicly accessible ASR system for Nigerian Pidgin - Free speech corpus for Nigerian Pidgin - First parallel (speech-to-text) Nigerian …