BiCycle: Group-wise Recursive Transformer Based on ASR Mechanism

Recursive transformer (RT) is a promising parameter-sharing technique for reducing computational burden of large-scale model. While RT has been successfully applied to large language models (LLMs), its effectiveness in automatic speech recognition (ASR) remains limited, despite the parallel trend of model scaling in the speech domain. In this paper, we reveal that conventional RT designs for LLMs are suboptimal for speech recognition, primarily because they do not fully consider the layer-wise specialization inherent in the ASR architecture, where lower layers focus on phonetic features and upper layers capture linguistic localization. To address this, we propose BiCycle, a novel RT scheme tailored for ASR.

In particular, we firstly analyze attention patterns in a pretrained ASR model to divide its layers into phonetic and linguistic groups. BiCycle then constructs an efficient RT model by transferring the pre-trained model's weights in a step-wise manner and applies recursion separately to the phonetic and linguistic groups, preventing conflicts between their roles. To further maximize BiCycle's performance, we propose groupwise feature distillation (GFD), which performs feature-level knowledge distillation (KD) between the teacher and student models' phonetic and linguistic groups, thereby effectively transferring the teacher model's knowledge tailored to each group's role. Extensive experimental results confirm that the proposed method not only preserves the original ASR mechanism but also outperforms conventional RT approaches.