基于谱时序图注意力池化和多任务学习的查询式关键词探测
计算与语言
2024-11-26 v2 人工智能
声音
音频与语音处理
摘要
现有的关键词探测（KWS）系统主要依赖预定义的关键词短语。然而，识别定制化关键词的能力对于与智能设备进行定制化交互至关重要。本文提出了一种新颖的查询式关键词探测（QbyE）KWS 系统，采用谱时序图注意力池化和多任务学习。该框架旨在有效学习 QbyE KWS 任务中 speaker-invariant 和 linguistic-informative 的嵌入。 Within this framework, we investigate three distinct network architectures for encoder modeling: LiCoNet, Conformer and ECAPA_TDNN. The experimental results on a substantial internal dataset of speakers have demonstrated the effectiveness of the proposed QbyE framework in maximizing the potential of simpler models such as LiCoNet. Particularly, LiCoNet, which is 13x more efficient, achieves comparable performance to the computationally intensive Conformer model (1.98% vs. 1.63% FRR at 0.3 FAs/Hr)。
引用
@article{arxiv.2409.00099,
title = {Query-by-Example Keyword Spotting Using Spectral-Temporal Graph Attentive Pooling and Multi-Task Learning},
author = {Zhenyu Wang and Shuyu Kong and Li Wan and Biqiao Zhang and Yiteng Huang and Mumin Jin and Ming Sun and Xin Lei and Zhaojun Yang},
journal= {arXiv preprint arXiv:2409.00099},
year = {2024}
}