基于特征流融合的带噪语音检测算法  

Noisy voice detection algorithm based on feature stream fusion

作　　者：龙华[1] 杨明亮 邵玉斌[1] LONG Hua;YANG Mingliang;SHAO Yubin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650031,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650031

出　　处：《通信学报》2020年第4期134-142,共9页Journal on Communications

基　　金：国家自然科学基金资助项目(No.61761025)。

摘　　要：针对语音通话中语音段的起始检测性能不佳,检测语音连续性结构受到破坏的问题,提出了一种基于特征流融合的带噪语音检测算法。首先,根据语音特性分别提取时域特征流、谱图特征流和统计特征流;其次,利用不同的语音特征流分别对带噪音频中的语音段进行概率估测;最后,将各个特征流估测得到的语音估测概率进行加权融合,并利用隐马尔可夫模型对语音估测概率进行短时状态处理。通过对复合语音数据库在多类型噪声与不同信噪比条件下的性能测试表明,所提算法相对于基于贝叶斯与DNN分类器的基线模型相比,语音检测正确率分别提高了21.26%与11.01%,显著提高了目标语音的质量。Aiming at the problem that the initial detection performance of voice segment was poor,and the voice continuity structure was damaged in voice communication,a noisy voice detection algorithm based on feature stream fusion was proposed.Firstly,the time domain feature stream,the spectral pattern feature stream and the statistical feature stream were extracted according to the voice characteristics.Secondly,the voice segment in the noisy audio was estimated by different voice feature streams.Finally,the voice prediction probability obtained by each feature stream was weighted and fused,and the voice estimation probability was processed in short time by the hidden Markov model.The performance test of composite voice database under the condition of multi-type noise and different signal-to-noise ratio shows that compared with the baseline model based on Bayesian and DNN classifier,the voice detection accuracy of the proposed algorithm is improved by 21.26%and 11.01%respectively,and the quality of target voice is significantly improved.

关 键 词：语音通话 语音检测 特征流融合 隐马尔可夫模型 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]