CONF VILLATORO-TELLO_ICASSP'24_2023/IDIAP Probability-Aware Word-Confusion-Network-to-Text Alignment Approach for Intent Classification Villatoro-Tello, Esaú Madikeri, Srikanth Sharma, Bidisha Khalil, Driss Kumar, Shashi Iuliia, Nigmatulina Motlicek, Petr Ganapathiraju, Aravind Cross-modal Alignment Intent Classification knowledge distillation Spoken Language Understanding Word-Confusion-Networks EXTERNAL https://publications.idiap.ch/attachments/papers/2024/VILLATORO-TELLO_ICASSP24_2023.pdf PUBLIC Proceedings of the 49th IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP) 2024 Seoul, Republic of Korea 2024 IEEE 12617-12621 https://ieeexplore.ieee.org/document/10445934 URL 10.1109/ICASSP48485.2024.10445934 doi Spoken Language Understanding (SLU) technologies have seen a big improvement due to the effective pretraining of speech representations. A common requirement of industry-based solutions is the portability to deploy SLU models in voice-assistant devices. Thus, distilling knowledge from large text-based language models has become an attractive solution for achieving good performance and guaranteeing portability. In this paper, we introduce a novel architecture that uses a cross-modal attention mechanism to extract bin-level contextual embeddings from a word-confusion network (WNC) encoding such that these can be directly compared and aligned with traditional text-based contextual embeddings. This alignment is achieved using a recently proposed tokenwise constrastive loss function. We validated our architecture's effectiveness by fine-tuning our WCN-based pretrained model to perform intent classification on the SLURP dataset. Obtained accuracy (81%), depicts a 9.4% relative improvement compared to a recent and equivalent E2E method.