合成语声的声学分析及识别特征算法
Acoustic analysis and recognition feature algorithm of synthetic speech
周峻林 1胡晓光 1黄子旭 1汪旭 1付哲宇1
作者信息
摘要
当前社会新型犯罪中电信诈骗案件频发,急需一种能够自动有效区分语声真伪的方法.为进一步增强目前深度学习领域识别合成语声的能力,为保障语声信息安全提供技术上的支持,针对合成语声声学特性上异于真实语声的特点,分析对比合成语声和真实语声的声学特性,设计了一种声学特征均方根角量化语声声强变化程度,结合基频变化率和语声窄带频谱图声学特征进行融合,量化了声学特性差异,聚焦了合成语声中关键声学信息.在神经网络模型中融合输入声学特征,在FoR数据集的验证集上得到了 0.6％的等错误率,在测试集上最好结果达到了 10.8％的等错误率.该文成功实现了对合成语声的识别,证实了声学特征的有效性和研究方案的可行性,在一定程度上拓宽了合成语声特征设计的研究思路.
Abstract
With the frequent occurrence of telecommunication fraud cases in the current new social crimes,a method that can automatically and effectively distinguish the authenticity of speech is urgently needed.To further enhance the current capability of detecting synthetic speech in the field of deep learning and to provide technical support for securing speech information,we analyze and compare the acoustic characteristics of synthetic speech and real speech,design an acoustic feature root mean square angle to quantify the variation of speech intensity,combine fundamental frequency variation and speech narrowband spectrogram acoustic features for fusion,quantify the difference of acoustic characteristics,and focus on the key acoustic information in synthetic speech.The fusion of input acoustic features in the neural network model yielded an equal error rate of 0.6％on the validation set of the FoR dataset,and the best result reached an equal error rate of 10.8％on the test set.The recognition of synthetic speech was successfully achieved,confirming the effectiveness of acoustic features and the feasibility of the research scheme of this paper,broadening the research ideas of synthetic speech feature design to a certain extent.
关键词
声学特征/声强/基频/语声频谱图/神经网络Key words
Acoustic features/Sound intensity/Fundamental frequency/Speech spectrogram/Neural network引用本文复制引用
基金项目
中国人民公安大学拔尖创新人才培养项目(2021)(2021yjsky017)
出版年
2024
NETL
NSTL
维普
万方数据