时频图像特征用于声场景分类

Time-frequency representation based feature extraction for audio scene classification

为解决根据音频流识别声场景的问题,对音频信号进行恒Q变换,得到其时频表达图像,然后进行滤波平滑等处理,随之提取能够表述信号谱能量变化方向信息的梯度直方图特征,以及能够捕捉信号谱纹理信息的局部二值模式特征,输入具有线性核函数的支持向量机分类器,对不同声场景数据进行分类实验。结果表明,相对于传统的时频域特征和梅尔频率倒谱系数特征,所提出的特征基本能够捕捉到给定声场景具有区分度的信息,所得分类率更高,且两者的互补作用使得联合特征分类效果达到最优,该方法为声信号特征提取贡献了一种新思路。

To recognize audio scene in a complex environment according to an audio stream, a constant-Q transform is chosen to obtain the timefrequency representation TFR of the signal. Due to the lack of prior knowledge on the signal and noise, a mean filtering is used to smooth the TFR image, then the features based on the histogram of gradients HOG of the TFR image are extracted, which can reflect the local direction of variation both in time and frequency of the signal power spectrum. Consequently the Local Binary Pattern LBP feature is considered, which captures the texture information of the signal. As for the classification algorithm, support vector machine with linear kernel function is used. Classification experiment has been done on the data of different acoustic scenes. Compared with the classical audio features such as MFCCs, the proposed features capture the discriminative power of a given audio scene to show good performance in classification, and the combined features achieve the best results. It is valuable in the field of feature extraction of acoustic signal.