摘要
低资源条件下的语音关键词检测是一个具有挑战性的问题,因为传统的基于大词汇量连续语音识别(LVCSR)的语音关键词检测方法不再适用.针对此问题提出了一种基于深度神经网络(DNN)输出层后验概率特征和改进的动态时间规整(DTW)算法的语音关键词检测方法.采用无监督高斯混合模型(GMM)和中、英文DNN音素模型得出的输入特征构建互补的子系统,并在SWS2013多语种数据集上进行实验.结果表明:相对于基线系统,分数层面的多语种、多系统融合能够有效地提升语音关键词检测系统的性能.
Spoken term detection in low-resource situations is a challenging task, because traditional large vocabu- lary continuous speech recognition (LVCSR)approaches are often unusable. We propose a query-by-example (QBE) spoken term detection (STD)method based on deep neural network (DNN)posteriorgram features and a modified dy- namic time warping (DTW) research approach. Subsystems are built with unsupervised Gaussian mixture model (GMM) and DNN monophone models trained on Chinese and English languages. The subsystems are then evaluated on the SWS2013 multilingual database of low-resource languages. The score-level fusion of these different languages and different subsystems is shown to improve performance significantly compared with the baseline results.
出处
《天津大学学报(自然科学与工程技术版)》
EI
CAS
CSCD
北大核心
2015年第9期757-760,共4页
Journal of Tianjin University:Science and Technology
基金
国家自然科学基金资助项目(61370034
61273268
61403224)
关键词
样例查询
语音关键词检测
DNN输出层特征
动态时间规整
query-by-example
spoken term detection
deep neural network output features
dynamic time warping