基于CNN的连续语音说话人声纹识别

发布时间：2019-02-13 07:28

【摘要】：近年来,随着社会生活水平的不断提高,人们对机器智能人声识别的要求越来越高。高斯混合—隐马尔可夫模型(Gaussian of mixture-hidden Markov model,GMM-HMM)是说话人识别研究领域中最重要的模型。由于该模型对大语音数据的建模能力不是很好,对噪声的顽健性也比较差,模型的发展遇到了瓶颈。为了解决该问题,研究者开始关注深度学习技术。引入了CNN深度学习模型研究连续语音说话人识别问题,并提出了CNN连续说话人识别(continuous speaker recognition of convolutional neural network,CSR-CNN)算法。模型提取固定长度、符合语序的语音片段,形成时间线上的有序语谱图,通过CNN提取特征序列,经过奖惩函数对特征序列组合进行连续测量。实验结果表明,CSR-CNN算法在连续—片段说话人识别领域取得了比GMM-HMM更好的识别效果。
[Abstract]:In recent years, with the continuous improvement of social living standards, the demand of machine intelligent voice recognition is becoming higher and higher. Gao Si Hybrid-Hidden Markov Model (Gaussian of mixture-hidden Markov model,GMM-HMM) is the most important model in the field of speaker recognition. Because the modeling ability of the model for large speech data is not very good, and the robustness to noise is also relatively poor, the development of the model has encountered a bottleneck. In order to solve this problem, researchers begin to pay attention to the technology of deep learning. In this paper, CNN depth learning model is introduced to study the continuous speech speaker recognition problem, and a CNN continuous speaker recognition (continuous speaker recognition of convolutional neural network,CSR-CNN) algorithm is proposed. The model extracts the speech fragments of fixed length and accords with the word order, and forms the ordered linguistic spectrum on the time line. The feature sequences are extracted by CNN, and the combination of feature sequences is continuously measured by the reward and punishment function. Experimental results show that the CSR-CNN algorithm achieves better recognition performance than GMM-HMM in the field of continuous-segment speaker recognition.
【作者单位】：杭州电子科技大学;
【分类号】：TP393

【相似文献】