ASR Model Adaptation with Domain Prompt Tuning

Automatic speech recognition (ASR) systems are already used in many industrial applications and, therefore, they need to be adapted to new domains with small memory and deployment overhead. In this work, we introduce Domain-Prompts, a methodology that involves training a small number of pluggable domain embedding parameters on a base ASR model for fast domain adaptation. On Aishell-1, Aishell-4, Ancient Poetry, and Children domains, the Domain-Prompts can achieve an absolute reduction of CER (Character Error Rate) in the range of 1.47%-11.73%. This improvement is comparable to the fully fine-tuned method, even though only 2.5% of the parameters of the base ASR model are updated. Additionally, our method is deployment-friendly as the learned domain embeddings are prefixed to the input of each model layer, rather than changing the structure and parameters of the base model. Therefore, our method is an ideal choice for on-the-fly adaptation of ASR systems, scaling them to new domains.