Parameter-efficient Fine-tuning of Conformer-based Streaming Speech Recognition into Non-streaming Models

In this paper, we propose a parameter-efficient fine-tuning method for converting streaming speech recognition models into non-streaming models. We introduce the Full-Context Adapter, a modified residual adapter designed to incorporate full-context information, enabling non-streaming conversion while keeping the pre-trained streaming model. The proposed method integrates the Full-Context Adapter into a Conformer-based RNN-Transducer streaming model, allowing it to operate as a non-streaming model. We conduct experiments to evaluate the effectiveness of our method, demonstrating that it achieves comparable performance to a non-streaming baseline model while requiring <10% additional parameters. Furthermore, we analyze optimal hyperparameter configurations and explore performance enhancement strategies. Our results show that the proposed approach provides a simple yet effective solution for efficiently operating both streaming and non-streaming ASR.