McLachlan N., Kumar D.

Recent studies on human recognition of environmental sounds have shown the importance of temporal information about signal amplitudes within octave band filter channels for people’s ability to recognize sound sources. This paper reports on the use of Fourier Transforms and Wavelet Transforms to produce spectral data of sounds from a variety of sources for classification by neural networks. It was found that the multi-resolution time-frequency analyses of wavelet transforms dramatically improved classification accuracy when statistical descriptors that captured measures of band limited spectral energy and temporal energy fluctuation were used. Ongoing research is focussing on the use of background subtraction techniques and source directionality information to improve the systems capacity to detect specific sound sources in complex acoustic fields.