Classification methods for Italian vowels

The automatic recognition of vowel sounds in voice signals is discussed. At a 10 to 12 kHz sampling rate for orthophonic signals, data are extracted using a 256 bit Hamming window. Two analysis methods are considered. The fast Fourier transform followed by cepstral analysis, and linear prediction which defines the signal as an autoregressive process of n order. The choice of parameters include the energy associated to spectrum bands, the first cepstral coefficients, the first four formant frequencies and the linear prediction coefficients or its derivatives. The stability problem and the elimination of multivowel sounds is analyzed.