Vowel and consonant recognition in Turkish using neural networks toward continuous speech recognition

This paper describes an artificial neural network that recognizes vowels and consonants in Turkish when a word or word sequence is uttered. The recognition is performed by three groups of connected nets. Inputs of the nets are the magnitudes of the short time spectrum at 16 mel-scaled frequency points in the range of 240 to 4500 Hz. The first net differentiates the input as being a vowel or a consonant. The outputs of this net as well as the magnitudes of the short time spectrum are used as inputs to the nets for vowel and consonant recognition. Consonant-vowel (CV) or vowel-consonant (VC) fragments of 11 speakers were used for training the different nets, and fragments from 7 more speakers were used for testing