基于深度学习的小样本声纹识别研究

Research on Small Sample Voiceprint Recognition Based on Deep Learning

摘要: 本文研究了小样本声纹识别问题。实验中采用梅尔倒谱系数与其动态差分系数组成的39维特征作为基本声学特征，再将基本声学特征通过由三层受限玻尔兹曼机堆叠而成的深度置信网络提取128维深度声学特征，最后通过支持向量机和随机森林进行分声纹识别。训练深度置信网络时，每个说话人选用短时语音信号组成的小样本数据作为该网络的训练集，同时将训练好的深度置信网络模型作为深度声学特征提取器，用该特征提取器对非训练集中说话人语音信号提取深度声学特征，进一步验证了该深度声学特征提取器的泛化能力。实验结果表明，本文设计的声纹识别模型识别准确率高，且深度特征提取器有较好的泛化能力。

Abstract: This paper studies the problem of small sample voiceprint recognition. In the experiment, the 39-dimensional features composed of the Mel cepstral coefficient and its dynamic differential coefficients are used as the basic acoustic features, and the basic acoustic features are extracted from the 128-dimensional depth acoustics through a deep belief network stacked by a three-layer restricted Boltzmann machine. Finally through the support vector machine and random forest for voiceprint recognition. Training deep belief networks, each speaker to choose short speech signal of small sample data as the network training set, trained deep belief network model at the same time as the depth of the acoustic feature extraction, with the characteristics of the extractor on the depth of the training focus on the speaker voice signal extraction acoustic characteristics, the generalization ability of the depth acoustic feature extractor is further verified. The experimental results show that the soundprint recognition model designed in this paper has high recognition accuracy, and the depth feature extractor has better generalization ability.