Domain Adaptation for ASR

Domain adaptation studies how to make speech recognizers work reliably when target speech differs from their training data, such as medical conversations, dialects, children’s speech, or noisy recordings. Common approaches include supervised fine-tuning and parameter-efficient methods such as LoRA and adapters, alongside synthetic or text-only training, pseudo-labeling, contextual prompting, and test-time adaptation. Researchers evaluate these methods on domain-specific speech benchmarks using word or character error rates, while also checking performance on other domains and speakers. A central tension is gaining accuracy with limited target data without losing generalization, increasing inference cost, or exposing private information.