通过对比微调与蒸馏实现轻量化且可泛化的声学场景表征
声音
2026-02-13 v2 机器学习
音频与语音处理
信号处理
摘要
声学场景分类（ASC）模型在边缘设备上通常 operate under fixed class assumptions, lacking the transferability needed for real-world applications that require adaptation to new or refined acoustic categories. 我们提出ContrastASC，通过结构化嵌入空间来保持场景之间的语义关系，从而实现对未见类别的适应，而无需重新训练。我们的 approach 结合了对预训练模型的监督对比微调和对比 representation distillation，将这种结构化知识传递给紧凑的 student 模型。我们的评估显示，ContrastASC 在适应未见类别方面表现出 improved few-shot adaptation,同时保持 strong closed-set performance.
关键词
引用
@article{arxiv.2510.03728,
title = {Lightweight and Generalizable Acoustic Scene Representations via Contrastive Fine-Tuning and Distillation},
author = {Kuang Yuan and Yang Gao and Xilin Li and Xinhao Mei and Syavosh Zadissa and Tarun Pruthi and Saeed Bagheri Sereshki},
journal= {arXiv preprint arXiv:2510.03728},
year = {2026}
}