A PC-based neural network for recognition of difficult syllables using LPC coefficient difference

The recognition of difficult CV (consonant-vowel) and VC (vowel-consonant) syllables using PC-based neural network and linear predictive coding (LPC) coefficients is studied. The speech corpus consisted of 41 syllables produced by three speakers in three different vowel contexts. The input to the neural network was differences in LPC coefficients sampled at three time instances in each syllable. Fully connected three layered backpropagation networks were trained by the delta learning rule. With relatively few parameters for each syllable, based on 123 tokens of 41 difficult syllables spoken with sentence context by three speakers (one male, two female), the preliminary results indicate that recognition accuracy is as high as 92.7%