Applicability of KanAdapter Beyond Speech Classification

Determine whether KanAdapter is effective for sequence-generation tasks such as automatic speech recognition and for foundation models in other modalities, including image, video, and large language models.

Background

The study evaluates KanAdapter on speech classification tasks—speaker verification, speech emotion recognition, and deepfake detection—and reports that its benefits vary across speech SSL backbones. Its effectiveness for sequence-generation tasks and for models outside the speech domain is not empirically established. The paper specifically identifies automatic speech recognition, other modalities, and LLMs as settings requiring further verification.

References

Third, our evaluation focuses on speech classification tasks (SV, SER, and DFD); applicability to sequence-generation tasks such as ASR, and to other modalities and LLMs, remains to be verified.

— KanAdapter: A Kolmogorov-Arnold Network-based Plug-and-Play Module for Efficient Fine-tuning of Foundation Speech Models  (2609.05281 - Dat et al., 4 Sep 2026) in Section Limitations