Application of Time‐Frequency Techniques to Sound Signals: Recognition and Diagnosis
Manuel Davy · 2008
Time-frequency representations are powerful analysis tools, especially in a non-stationary context. They can be employed for the classification or detection of signals, in the extremely frequent situation where we have some training signals but no mathematical model of the signal to be processed. We then speak of supervised non-parametric decision. In this chapter, we propose two applications involving sound signals: the verification of speakers and the detection of defects in acoustic loudspeakers. In the approach that we propose, time-frequency signal representations are used as basic information for constructing a decision-making process. Two time-frequency representations are compared using specially designed distance measures. During the training phase, the kernel of the time-frequency representations as well as the distance measure are optimized so as to minimize the probability of error. The latter is estimated for the whole set of training signals.