Hybrid Architectures for Low-Resource Speech Recognition: Integrating End-to-End Models with Graph-Based Language Models

Abstract

Low-resource speech recognition systems face significant challenges due to limited training data and computational resources. This paper proposes a novel hybrid architecture that integrates end-to-end speech recognition models with graph-based language models to improve performance in low-resource settings. The hybrid approach combines the strengths of deep neural networks and language models to enhance recognition accuracy and robustness. We evaluate the proposed system on multiple low-resource languages and compare its performance with traditional speech recognition models.