The project compares the wavelet and cepstrum transforms for recognizing plosive and vocalic phonemes, using energy level coefficients for wavelet and the 14 first coefficients but the first for cepstrum. As a classifier a similar neural network was chosen for both systems. The chosen phonemes were /p/, /t/, /k/, /a/ and /e/, taken from a similar number of sample windows from a database of 93 phrases pronounced by a single speaker. Wavelet and cepstrum had a similar performance for recognizing plosive phonemes. For vocalic phonemes cepstrum had definitively better results.
