Sound Detection through Transient Models using Wavelet Coefficient Trees

Michel Vacher (1) , Dan Istrate (1) , Jean-François Serignat (1)

1 CLIPS-IMAG - Equipe GEOD, Groupe d'étude sur l'oral et le dialogue

Medical Telesurveillance needs human operator to be as-sisted by smart information systems. Therefore automatic determination of sound type emitted in patient's habita-tion may greatly increase the versatility of such a system. Sounds are acquired through microphones set out in each room. Detection is the first step of our sound analysis sys-tem and is necessary to extract the significant sounds before initiating the classification step. This paper proposes a de-tection method using transient models, based upon dyadic trees of wavelet coefficients to insure short detection delay. This method is used to detect at once the beginning and the end of the audio signal allowing signal extraction in noisy environment. The precision of this step is important to avoid a decrease of performances during the second step which is the classification step. This step uses a Gaussian Mixture Model classifier with classical acoustical param-eters like MFCC. Detection and classification stages are evaluated in experimental recorded noise condition which is non-stationary and more realistic than simulated white noise. Wavelet filtering methods are proposed to enhance classification performances in low signal to noise ratios.

Mots clés: Noise; Sound Extraction; Sound Classification; Wavelet Transform

Domaines: Autre [cs.OH]

Michel Vacher, Dan Istrate, Jean-François Serignat. Sound Detection through Transient Models using Wavelet Coefficient Trees. Complex Systems, Intelligence and Modern Technology Applications, Sep 2004, Cherbourg, France. pp.367-372. ⟨hal-01088260⟩