Based on the spectrograms analysis and the individual frequency bands of speech under G-force, in this pa-per, a new Mel frequency scale is proposed, and the related MFCC (Mel Frequency Cepstrum Coefficient) is adopte...Based on the spectrograms analysis and the individual frequency bands of speech under G-force, in this pa-per, a new Mel frequency scale is proposed, and the related MFCC (Mel Frequency Cepstrum Coefficient) is adoptedas the features for recognition of stressed speech under G-force. It is shown from the experiments that the proposedmethod is better than other methods of Mel-based features for stressed speech recognition.展开更多
为提高水下蛙人呼吸声识别的准确度,提出一种基于Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)的蛙人呼吸声信号特征匹配方法。计算呼吸声信号之间、信号与环境噪声及舰船辐射噪声的MFCC夹角和MFCC距离并进行匹配比较,以...为提高水下蛙人呼吸声识别的准确度,提出一种基于Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)的蛙人呼吸声信号特征匹配方法。计算呼吸声信号之间、信号与环境噪声及舰船辐射噪声的MFCC夹角和MFCC距离并进行匹配比较,以进行分类识别。某湖试验数据的处理结果表明:蛙人呼吸声与舰船辐射噪声及环境噪声的MFCC参数有着明显的差异,能够对蛙人呼吸声信号与干扰噪声进行区分,证明了基于MFCC特征算法的有效性,对发展港口、码头等近海海域附近的水下蛙人探测声呐和预警系统具有实际意义。展开更多
文摘Based on the spectrograms analysis and the individual frequency bands of speech under G-force, in this pa-per, a new Mel frequency scale is proposed, and the related MFCC (Mel Frequency Cepstrum Coefficient) is adoptedas the features for recognition of stressed speech under G-force. It is shown from the experiments that the proposedmethod is better than other methods of Mel-based features for stressed speech recognition.
文摘为提高水下蛙人呼吸声识别的准确度,提出一种基于Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)的蛙人呼吸声信号特征匹配方法。计算呼吸声信号之间、信号与环境噪声及舰船辐射噪声的MFCC夹角和MFCC距离并进行匹配比较,以进行分类识别。某湖试验数据的处理结果表明:蛙人呼吸声与舰船辐射噪声及环境噪声的MFCC参数有着明显的差异,能够对蛙人呼吸声信号与干扰噪声进行区分,证明了基于MFCC特征算法的有效性,对发展港口、码头等近海海域附近的水下蛙人探测声呐和预警系统具有实际意义。
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%。
