Image processing for time-frequency speech analysis

摘要

Classic notion of spectral analysis is based on stationary hypothesis. But, stationary signals constitute only an exception when we consider the immense majority of real signals found in nature.rnTime-frequency representations and particularly those derived from the Wigner-Ville transformation is well adapted to non-stationary signal analysis. However, the resulting unwanted cross terms, in the time-frequency plane, make it very difficult for anyone to draw straightforward conclusions.rnAttenuation of these terms requires a smoothing depends on the kernel choice characterizing the representations which have various properties and performances. These properties are still not very known.rnIn order to bypass this problem, image processing techniques are applied to the time-frequency representations. The method is based on image binarisation which separates the object from the background. The application to speech signal as a non stationary signal and a communication tool, to locate the mains frequencies of the French vowels has given very satisfactory results.