Attention connectivity: An interpretable framework for revealing salient long-range connectivity in self-supervised speech transformers

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Self-supervised speech Transformers have achieved strong performance in automatic speech recognition; however, their cross-layer information organization remains poorly understood, limiting explainability. While existing explainability studies have examined individual attention patterns, head characteristics, and token-level cross-layer attribution, model-wide statistical dependencies among attention heads across layers remain insufficiently characterized. Inspired by dynamic functional connectivity analysis in neuroscience, we propose Attention Connectivity, an interpretable framework that quantifies dependencies among attention heads across layers to characterize interaction patterns. Through this framework, we identify Salient Long-range Connectivity, a structural motif spanning multiple layers. This motif demonstrates heightened sensitivity, directional dependency characteristics, and non-redundant functional contributions; notably, pruning analyses provide functional evidence that the associated attention heads make important contributions to recognition performance. Furthermore, acoustic alignment analysis shows that these connections are consistently associated with energy-related characteristics of continuous speech. Experiments across wav2vec 2.0 variants, HuBERT, WavLM, and an additional speech dataset reveal recurring salient cross-layer connectivity patterns, supporting the robustness and cross-model generalizability of the proposed analysis. This work provides a systematic methodology for uncovering structured cross-layer organization in speech Transformers and advances connectivity-based attention explainability.