一种基于噪声场景识别与多特征集成学习的活动语音检测方法

An Approach for Voice Activity Detection Based on Noise Recognition and Multi-feature Ensemble Learning

为了提高动态噪声环境下语音检测的准确性,提出一种基于噪声场景识别与多特征集成学习的活动语音检测方法。针对非平稳复杂音频信号,提取小波能量、奇异值等时频域特征,采用随机森林优选出可分性更好的特征组合。实验表明,相比于支持向量机与多层感知器分类模型,所提方法的语音检测准确率提升显著,并且在不同噪声类型、不同噪声强度下性能更加稳定。

To improve the accuracy of voice activity detection under dynamic noise environments,this paper proposes an approach based on noise recognition and multi-feature ensemble learning.Considering the nonstationary and complex signals,time-frequency features,like wavelet energies and singular values,are extracted and selected using Random Forest(RF)for better separability.Case studies based on different noises and different noise powers demonstrate the higher accuracy and stability of the proposed approach.