A Framework for End-to-End Speech Recognition using LLM-Generated Synthetic Data for Contextual Biasing

Automatic Speech Recognition (ASR) systems often exhibit performance degradation when encountering domain-specific terminology, named entities, or other contextually dependent expressions not well-represented in general-purpose training corpora. Conventional domain adaptation techniques typically require substantial in-domain speech-text paired data, which is often scarce, expensive to acquire, or restricted by privacy concerns. This paper proposes a novel framework to address this challenge by leveraging Large Language Models (LLMs) and Controllable Speech Synthesis (CSS) to generate a synthetic, context-aware corpus for ASR model adaptation. Our approach operates in three stages: (1) An LLM is guided by a structured prompting strategy to generate diverse, domain-specific text. (2) A CSS model synthesizes this text into a rich and acoustically varied speech corpus, simulating diverse speaker characteristics, prosodic patterns, and environmental conditions. (3) The resulting synthetic speech-text pairs are used to fine-tune a pre-trained ASR model, enhancing its contextual awareness. This method requires no pre-existing in-domain data. Experimental results on multiple datasets demonstrate that our framework achieves a relative Character Error Rate (CER) reduction of 6–12% on context-specific phrases while preserving performance on general-domain speech. This work demonstrates that LLM-driven synthetic data generation provides a scalable and effective solution for mitigating contextual bias in ASR systems, particularly in low-resource settings.