摘要
针对现有语音关键词检测方法定位精度低的问题,提出了一种基于多尺度距离矩阵的语音关键词检测与细粒度定位方法(spoken term detection and fine-grained localization method based on multi-scale distance matrices,MF-STD)。该方法首先利用残差卷积网络提取特征并构建距离矩阵以建模输入之间的相关性;其次通过多尺度分割和解耦头学习不同尺度下的定位信息;最后根据多尺度加权定位损失、置信度损失和分类损失优化模型,实现对关键词存在性和时域边界的细粒度预测。在LibriSpeech数据集上的实验结果表明,MF-STD在集内词的检测中,精准率和交并比分别达到97.1%和88.6%;在集外词的检测中,精准率和交并比分别达到96.7%和88.2%。与现有的语音关键词检测与定位方法相比,MF-STD的检测准确率和定位精度显著提升,充分证明该方法的先进性,也证明了多尺度特征建模与细粒度定位约束在语音关键词检测任务中的有效性。
Aiming to address the low localization accuracy of existing spoken term detection methods,this paper proposed a spoken term detection and fine-grained localization method based on multi-scale distance matrices(MF-STD).This method firstly employed a residual convolutional network to extract features and construct a distance matrix to model the correlation between inputs.Then,it learnt the localization information at different scales through multi-scale segmentation and decoupling heads.Finally,the model was optimized according to the multi-scale weighted localization loss,confidence loss,and classification loss.This enabled the model to achieve fine-grained prediction of keyword existence and time domain boundaries.Experimental results on the LibriSpeech dataset demonstrate that for in-vocabulary detection,the precision and intersection over union(IoU)reach 97.1%and 88.6%,respectively.In the case of out-of-vocabulary detection,the precision and IoU reach 96.7%and 88.2%,respectively.In comparison to existing methods for spoken term detection and localization,MF-STD significantly improves detection accuracy and localization precision.This fully demonstrates the superiority of the proposed method and the effectiveness of multi-scale feature modeling and fine-grained localization constraints in spoken term detection tasks.
作者
李祥瑞
毛启容
Li Xiangrui;Mao Qirong(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang Jiangsu 212013,China;Jiangsu Province Big Data Ubiquitous Perception&Intelligent Agriculture Application Engineering Research Center,Zhenjiang Jiangsu 212013,China)
出处
《计算机应用研究》
CSCD
北大核心
2024年第11期3370-3375,共6页
Application Research of Computers
基金
江苏省重点研发计划资助项目(BE2020036)
江苏大学应急管理学院专项科研项目(KY-A-01)。
关键词
语音关键词检测
语音细粒度定位
多尺度检测
残差卷积网络
spoken term detection
speech fine-grained localization
multi-scale detection
convolutional residual network
