长时语音特征在说话人识别技术上的应用

本文除介绍常用的说话人识别技术外,主要论述了一种基于长时时频特征的说话人识别方法,对输入的语音首先进行VAD处理,得到干净的语音后,对其提取基本时频特征。在每一语音单元内把基频、共振峰、谐波等时频特征的轨迹用Legendre多项式拟合的方法提取出主要的拟合参数,再利用HLDA的技术进行特征降维,用高斯混合模型的均值超向量表示每句话音时频特征的统计信息。在NIST06说话人1side-1side说话人测试集中,取得了18.7%的等错率,与传统的基于MFCC特征的说话人系统进行融合,等错率从4.9%下降到了4.6%,获得了6%的相对等错率下降。

Long span prosodic features for speaker recognition

In this paper, we first give an introduction about speaker recognition techniques. Then a novel speaker verification method based on long span prosodic features is proposed. After speech is pre-processed by a voice activity detection module, and basic prosody features are extracted for each speech unit, we carried out an approximation of the pitch, formant, time domain energy and harmonic energy contours by taking the leading terms in a Legendre polynomial expansion. HLDA is used to reduce the feature dimension and mean supervector in each individual Gaussian is used to represent the distribution of the time-frequency features. Experiments on NIST06 show that the proposed method can reduce the EER from 4.9% to 4.6% when fusing with the traditional MFCC-featured system.