基于LFBank与FBank混合特征的声纹识别研究

Study on Voiceprint Recognition Based on Mixed Features of LFBank and FBank

语音特征提取是声纹识别过程中的重要步骤,对于声音频率的分布男性与女性差距较大,但现有的特征提取算法并没有针对不同性别声音频率特性做出相应改进。针对上述问题,提出了为女性声纹识别所设计的语音特征提取算法LFBank,将线性滤波器组用于特征提取过程,利用其线性分布的特点弥补传统梅尔滤波器组提取高频区域信息时的不足。另一方面,为了突破单一性别局限,拓宽应用场景,综合线性滤波器组与梅尔滤波器组的优势,将LFBank与FBank特征结合得到混合特征向量进行声纹识别。将LFBank和常用特征FBank与MFCC进行实验对比,实验结果表明,基于线性滤波器组的特征向量在识别女性声音时更有优势。对于混合特征而言,在与单一特征的对比实验中,混合特征能够达到比单一特征更好的识别效果,具有更广泛的应用场景。

Speech feature extraction is an important step in the process of voiceprint recognition.There is a large gap between men and women in the distribution of sound frequency,but the existing feature extraction algorithms have not made corresponding improvements for the sound frequency characteristics of different genders.To solve the above problems,a speech feature extraction algorithm LFBank designed for female voiceprint recognition is proposed.The linear filter banks is introduced into the feature extraction process,and its linear distribution is used to make up for the deficiency of the traditional Mel filter banks in extracting high-frequency region information.On the other hand,in order to break through the limitation of single gender and broaden the application scenarios,combining the advantages of linear filter banks and Mel filter banks,LFBank and FBank features are combined to obtain mixed feature vectors for voiceprint recognition.The LFBank is compared with the commonly used feature FBank and MFCC,and experimental results show that the feature vector based on linear filter bank has more advantages in recognizing female voice.For mixed features,in the comparison experiment with single features,they can achieve better recognition effect than single features and have a wider range of application scenarios.