为了提高低信噪比下语种识别的准确率,引入一种新的特征提取融合方法.在前端加入有声段检测,并基于人耳听觉感知模型提取伽玛通频率倒谱系数(Gammatone Frequency Cepstrum Coefficient,GFCC)特征,通过主成分分析对特征进行压缩、降噪,...为了提高低信噪比下语种识别的准确率,引入一种新的特征提取融合方法.在前端加入有声段检测,并基于人耳听觉感知模型提取伽玛通频率倒谱系数(Gammatone Frequency Cepstrum Coefficient,GFCC)特征,通过主成分分析对特征进行压缩、降噪,融合每个有声段的Teager能量算子倒谱参数,通过高斯混合通用背景模型进行语种识别验证.实验结果表明,在信噪比为-5~0 dB时,相对于基于对数梅尔尺度滤波器组能量特征方法,融合特征集方法对5种语言的识别率,分别提升了23.7%~34.0%,其他信噪比等级下识别率也有明显的提升.展开更多
针对法庭说话人识别中待鉴定人员语音样本不足的问题,提出了一种新的对说话人自身变化性建模的替代性方法以及相应的方差控制算法。使用同条件下的参考数据库构建识别系统的多个相同说话人得分模型,代替检验需要的多个非同期的带检验人...针对法庭说话人识别中待鉴定人员语音样本不足的问题,提出了一种新的对说话人自身变化性建模的替代性方法以及相应的方差控制算法。使用同条件下的参考数据库构建识别系统的多个相同说话人得分模型,代替检验需要的多个非同期的带检验人员语音样本比较时的得分模型,以获得能反映说话人自身变化性的统计模型。基于目前最新的法庭证据评估的似然比证据强度评估体系,使用MFCC(Mel Frequency Cepstral Coefficients)和GFCC(Gammatone Frequency Cepstral Coefficients)特征对该方法的有效性进行了验证,并对上述特征进行了特征级和决策级融合。实验结果表明:该方法在纯净语音环境和噪声环境下都具有很高的识别率和稳定性,并且特征级融合能进一步提高识别系统的性能。展开更多
针对传统语音端点检测方法在噪声环境下鲁棒性较差以及对语音段检测效果不佳的问题,提出一种多特征融合的语音端点检测方法.首先,提取带噪语音信号的子带谱熵特征和基于Mel频率倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)的投...针对传统语音端点检测方法在噪声环境下鲁棒性较差以及对语音段检测效果不佳的问题,提出一种多特征融合的语音端点检测方法.首先,提取带噪语音信号的子带谱熵特征和基于Mel频率倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)的投影特征,并将Gammatone频率倒谱系数的第一维系数GFCC0特征应用到语音端点检测任务中;然后,对3类特征进行自适应加权融合,得到适用于端点检测的融合特征;最后,采用模糊C均值聚类自适应估计门限阈值,再通过双门限法得到端点检测的结果.所提方法和已有传统方法相比,在7种噪声环境下均取得了更好的端点检测结果,提升了语音端点检测的准确率,特别是在volvo噪声环境下的端点检测准确率可以达到94.5%以上.展开更多
The problem of disguised voice recognition based on deep belief networks is studied. A hybrid feature extraction algorithm based on formants, Gammatone frequency cepstrum coefficients(GFCC) and their different coeffic...The problem of disguised voice recognition based on deep belief networks is studied. A hybrid feature extraction algorithm based on formants, Gammatone frequency cepstrum coefficients(GFCC) and their different coefficients is proposed to extract more discriminative speaker features from the original voice data. Using mixed features as the input of the model, a masquerade voice library is constructed. A masquerade voice recognition model based on a depth belief network is proposed. A dropout strategy is introduced to prevent overfitting, which effectively solves the problems of traditional Gaussian mixture models, such as insufficient modeling ability and low discrimination. Experimental results show that the proposed disguised voice recognition method can better fit the feature distribution, and significantly improve the classification effect and recognition rate.展开更多
Identification of bird species from their sounds has become an important area in biodiversity-related research due to the relative ease of capturing bird sounds in the commonly challenging habitat. Audio features have...Identification of bird species from their sounds has become an important area in biodiversity-related research due to the relative ease of capturing bird sounds in the commonly challenging habitat. Audio features have a massive impact on the classification task since they are the fundamental elements used to differentiate classes. As such, the extraction of informative properties of the data is a crucial stage of any classification-based application. Therefore, it is vital to identify the most significant feature to represent the actual bird sounds. In this paper, we propose a novel feature that can advance classification accuracy with modified features, which are most suitable for classifying birds from its audio sounds. Modified Gammatone frequency cepstral coefficient(GTCC) features have been extracted with their frequency banks adjusted to suit bird sounds. The features are then used to train and test a support vector machine(SVM) classifier. It has been shown that the modified GTCC features are able to give 86% accuracy with twenty Bornean birds. Furthermore, in this paper, we are proposing a novel probability enhanced entropy(PEE) feature, which, when combined with the modified GTCC features, is able to improve accuracy further to 89.5%. These results are significant as the relatively low-resource intensive SVM with the proposed modified GTCC, and the proposed novel PEE feature can be implemented in a real-time system to assist researchers,scientists, conservationists, and even eco-tourists in identifying bird species in the dense forest.展开更多
文摘为了提高低信噪比下语种识别的准确率,引入一种新的特征提取融合方法.在前端加入有声段检测,并基于人耳听觉感知模型提取伽玛通频率倒谱系数(Gammatone Frequency Cepstrum Coefficient,GFCC)特征,通过主成分分析对特征进行压缩、降噪,融合每个有声段的Teager能量算子倒谱参数,通过高斯混合通用背景模型进行语种识别验证.实验结果表明,在信噪比为-5~0 dB时,相对于基于对数梅尔尺度滤波器组能量特征方法,融合特征集方法对5种语言的识别率,分别提升了23.7%~34.0%,其他信噪比等级下识别率也有明显的提升.
文摘为了解决传统径向基(Radial basis function,RBF)神经网络在语音识别任务中基函数中心值和半径随机初始化的问题,从人脑对语音感知的分层处理机理出发,提出利用大量无标签数据初始化网络参数的无监督预训练方式代替传统随机初始化方法,使用深度自编码网络作为语音识别的声学模型,分析梅尔频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)和基于Gammatone听觉滤波器频率倒谱系数(Gammatone Frequency Cepstrum Coefficient,GFCC)下非特定人小词汇量孤立词的抗噪性能。实验结果表明,深度自编码网络在MFCC特征下较径向基神经网络表现出更优越的抗噪性能;而与经典的MFCC特征相比,GFCC特征在深度自编码网络下平均识别率相对提升1.87%。
文摘针对法庭说话人识别中待鉴定人员语音样本不足的问题,提出了一种新的对说话人自身变化性建模的替代性方法以及相应的方差控制算法。使用同条件下的参考数据库构建识别系统的多个相同说话人得分模型,代替检验需要的多个非同期的带检验人员语音样本比较时的得分模型,以获得能反映说话人自身变化性的统计模型。基于目前最新的法庭证据评估的似然比证据强度评估体系,使用MFCC(Mel Frequency Cepstral Coefficients)和GFCC(Gammatone Frequency Cepstral Coefficients)特征对该方法的有效性进行了验证,并对上述特征进行了特征级和决策级融合。实验结果表明:该方法在纯净语音环境和噪声环境下都具有很高的识别率和稳定性,并且特征级融合能进一步提高识别系统的性能。
文摘针对传统语音端点检测方法在噪声环境下鲁棒性较差以及对语音段检测效果不佳的问题,提出一种多特征融合的语音端点检测方法.首先,提取带噪语音信号的子带谱熵特征和基于Mel频率倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)的投影特征,并将Gammatone频率倒谱系数的第一维系数GFCC0特征应用到语音端点检测任务中;然后,对3类特征进行自适应加权融合,得到适用于端点检测的融合特征;最后,采用模糊C均值聚类自适应估计门限阈值,再通过双门限法得到端点检测的结果.所提方法和已有传统方法相比,在7种噪声环境下均取得了更好的端点检测结果,提升了语音端点检测的准确率,特别是在volvo噪声环境下的端点检测准确率可以达到94.5%以上.
基金supported by Natural Science Foundation of Liaoning Province (Nos. 2019-ZD-0168 and 2020-KF-12-11)Major Training Program of Criminal Investigation Police University of China (No. 3242019010)+1 种基金Key Research and Development Projects of Ministry of Science and Technology (No. 2017YFC0821005)Second Batch of New Engineering Research and Practice Projects(No. E-AQGABQ20202710)。
文摘The problem of disguised voice recognition based on deep belief networks is studied. A hybrid feature extraction algorithm based on formants, Gammatone frequency cepstrum coefficients(GFCC) and their different coefficients is proposed to extract more discriminative speaker features from the original voice data. Using mixed features as the input of the model, a masquerade voice library is constructed. A masquerade voice recognition model based on a depth belief network is proposed. A dropout strategy is introduced to prevent overfitting, which effectively solves the problems of traditional Gaussian mixture models, such as insufficient modeling ability and low discrimination. Experimental results show that the proposed disguised voice recognition method can better fit the feature distribution, and significantly improve the classification effect and recognition rate.
文摘Identification of bird species from their sounds has become an important area in biodiversity-related research due to the relative ease of capturing bird sounds in the commonly challenging habitat. Audio features have a massive impact on the classification task since they are the fundamental elements used to differentiate classes. As such, the extraction of informative properties of the data is a crucial stage of any classification-based application. Therefore, it is vital to identify the most significant feature to represent the actual bird sounds. In this paper, we propose a novel feature that can advance classification accuracy with modified features, which are most suitable for classifying birds from its audio sounds. Modified Gammatone frequency cepstral coefficient(GTCC) features have been extracted with their frequency banks adjusted to suit bird sounds. The features are then used to train and test a support vector machine(SVM) classifier. It has been shown that the modified GTCC features are able to give 86% accuracy with twenty Bornean birds. Furthermore, in this paper, we are proposing a novel probability enhanced entropy(PEE) feature, which, when combined with the modified GTCC features, is able to improve accuracy further to 89.5%. These results are significant as the relatively low-resource intensive SVM with the proposed modified GTCC, and the proposed novel PEE feature can be implemented in a real-time system to assist researchers,scientists, conservationists, and even eco-tourists in identifying bird species in the dense forest.
