| This dissertation presents a new approach to acoustic signature recognition using a wavelet preprocessing. The new approach identifies the source of the sound waveform generated. The new approach uses a complete wavelet packet decomposition instead of a logarithmic wavelet tree. After the strengths of all the subbands are obtained, it is possible to distinguish one sound source from another by identifying the subbands that have significantly higher strengths than neighboring subbands. Although this new approach works well with machinery and bird sounds, it does not distinguish one human speaker from another. The main reason why it cannot distinguish speakers is that discrete wavelet transform is not time-shift invariant, which is important in detection problems. That is, the coefficients of a delayed signal are not the shifted version of the coefficients of the original signal. |
