Classification of acoustic noise signals using wavelet spectrum based support vector machine

Harsh noises come from air-conditioning units are chronic complaining issues to their users. Individual perceptions of noise levels have been generally quantified by means of subjective evaluation such as a jury test. This article proposes a classification approach to acoustic noise signals using a wavelet spectrum analysis. We derive energy spectrums of noise signals using a discrete wavelet transform at pre-specified window length. The energy spectrums are a linear form and represented by a Hurst parameter as an informative summary of long-range dependent signal data. The Hurst parameter controls the self-similarity scaling as well as the degree of long-range dependence. We estimate the Hurst parameter through the least squares regression of sample energy against a resolution level in the wavelet spectral domain. In the context of multi-class classification problem, the classification of noise signals is performed by a nonlinear support vector machine (SVM) for parameter estimates of linear energy profiles containing the Hurst parameter. In an application example of air-conditioner noise signals, empirical results show that the proposed method offers the higher level of accuracy in acoustic noise sound classification.