基于稀疏表示权重张量的音频特征提取算法

2016-05-14 10:37:23林静杨继臣张雪源李新超
计算机应用 2016年5期

林静 杨继臣 张雪源 李新超

摘要:为了更好地描述非平稳音频信号的特征,提出了一种基于Gabor字典和稀疏表示权重张量的时频音频特征提取方法。该方法基于Gabor字典将音频信号编码为稀疏的权重向量,并进一步将权重向量中的元素重新排列为张量形式,该张量各阶分别刻画了信号的时间、频率以及时长特性,为信号的联合时频长表示。通过对该张量进行因子分解,将分解后得到的频率因子和时长因子拼接为音频特征。针对稀疏张量分解时容易产生过拟合的问题,提出一种自调整惩罚参数分解算法并进行了改进。实验结果显示,所提出的特征相对于传统梅尔倒谱系数(MFCC)特征、MFCC特征及匹配追踪算法(MP)求解的特征联合拼接得到的MFCC+MP特征和非均匀尺度频率图特征对15类音效分类效果分别提升了28.0%、19.8%和6.7%。

关键词:稀疏表示;张量因子分解;音效分类;时频特征

中图分类号:TN912.3 文献标志码:A

Abstract:A joint timefrequency audio feature extraction algorithm based on Gabor dictionary and weight tensor of sparse representation was proposed to describe the characteristic of nonstationary audio signal. Conventional sparse representation uses a predefined dictionary to encode the audio signal as sparse weight vector. In this paper, the elements in the weight vector were reorganized into tensor format. Each order of the tensor respectively characterized time, frequency and duration property of signal, making it the joint timefrequencyduration representation of the signal. The frequency factors and duration factors were concatenated as audio features through tensor decomposition. To solve the overfitting problem of sparse tensor factorization, an automaticadjustpenaltycoefficient factorization algorithm was proposed. The experimental results show that the proposed feature outperforms MFCC (MelFrequency Cepstrum Coefficient) feature, MFCC+MP feature concatenated by MFCC and Matching Pursuit (MP) features, and nonuniform scalefrequency map feature by 28.0%, 19.8% and 6.7% respectively, in 15category audio classification.

Key words:sparse representation;tensor factorization;audio effect classification;timefrequency feature

0 引言

传统的语音特征通常从频域或者倒谱域对信号进行表示,这些特征均假定在一短时帧内信号的统计特性平稳,即信号具有短时平稳性。但是这些特征不适用于刻画非平稳的音频信号,例如闪电声和枪声。为了描述统计特性随时间变化的信号,基于稀疏表示(Sparse Representation)的方法受到了广泛的关注。稀疏表示利用一个过完备的字典,将信号编码为一个稀疏的权重向量序列,每个向量中的非零元素值表示了重建时对应原子的权重。有许多文献使用该权重向量进行信号分类。……

登录APP查看全文