语音活动检测对方言辨识系统的影响研究-《计算机技术与发展》

文章信息/Info

Title:: Study on Influence of Voice Activity Detection for Dialects Identification System

Author(s):: ZHANG Ning ; GU Ming-liang; ZHU Jun-mei; ZHOU Jie; [1]School of Physics ＆ Electronic Engineering, Jiangsu Normal University[2]School of Linguistic Science, Jiangsu Normal University

Keywords:: dialect identification; voice activity detection; DD＋Hang-over algorittun

摘要:: 分别把基于阈值判断和基于统计模型的语音活动检测（VAD）应用于汉语方言辨识系统中，对比了系统识别率及运算时间。其中基于能量、过零率等阈值判断的方法以其算法简单、计算量少的优点在高信噪比噪声环境下取得较好的效果，但在低信噪比噪声环境下准确性及鲁棒性急剧下降。在相同测试环境下，采用统计模型的DD＋Hang-over算法取代传统经典阈值算法。实验表明，基于统计模型的算法在高斯混合模型（GMM）系统下运算时间稍氏，但抗噪声性能明显优越于基于阈值判断的算法，尤其在低信噪比的加性噪声环境下效果更显著

Abstract:: Voice active detection （VAD） based on threshold judgement and statistical-model is used in Chinese dialect identification sys tem respectively. The recognition accuracy and computing time have also be compared. Threshold-based methods such as short term energy has achieved good results in a high SNR noisy environment with its simple algorithm and less computation. Wbereas,its performance declines sharply in the low SNR environment. Statistical model-based method with DD＋ Hang-over is used instead of the traditional threshold-based methods under the same test environment. Experiments with the system of GMM show that although statistical model. based method with DD＋Hang-over takes more time,it has better anti-noise performance, especially in the low SNR environment