Comparison of some time-frequency analysis methods for classification of plosives

The paper deals with the context independent recognition of unvoiced plosives (/p/, /t/, /k/). In several solutions the best feature vectors are being sought in the burst segments of plosives. It has been proved that the difference between stops fade out very quickly after the burst onset. The short time of the burst duration implies the need of the higher time resolution in time-frequency analysis. The paper presents the results of the application of selected methods of high resolution time-frequency distributions for the recognition of stops. Apart from the traditional Short Time Fourier Transform based spectrogram, Gabor Spectrogram and cone shaped distribution have been used to calculate input parameters (cepstral coefficients) to the neural network used for classification.