Spaces:
Running on Zero
Apply for a GPU community grant: Academic project
Project Description
I am a final-year Computer Science student at the University of Zambia conducting research on Low-Resource Automatic Speech Recognition for Zambian Languages. The project focuses on adapting and evaluating pretrained speech recognition models for Zambian languages, specifically Bemba, Nyanja, and Tonga.
The research investigates how transfer learning can be used to develop domain-specific ASR models for languages with limited speech resources. I am experimenting with pretrained models including Whisper, MMS, and XLS-R, fine-tuning them on Zambian language speech datasets, and comparing their performance using standard ASR evaluation metrics such as Word Error Rate (WER).
GPU access is essential for this research because fine-tuning modern speech models requires significant computational resources. The available computing resources locally are limited, making it difficult to efficiently train larger models, perform multiple experiments, and conduct comprehensive comparisons between different architectures and training approaches.
With Hugging Face GPU resources, I will be able to:
Fine-tune and evaluate larger pretrained ASR models.
Run controlled experiments across multiple Zambian languages.
Perform hyperparameter optimization and model evaluation faster.
Improve the quality of open-source speech recognition models and datasets.
Produce reproducible research outcomes as part of my final-year Computer Science project.
This project contributes toward addressing the lack of speech technology resources for African languages by developing and sharing models and datasets that can support future research in low-resource language technologies.