Difficult syllable recognition using LPC coefficient differences and PC-based neural network
C. Shim, Blas Espinoza-Varas, John Y. Cheung · 2002
An investigation was conducted of the recognition of difficult CV (consonant-vowel) syllables using PC-based neural network paradigms with LPC coefficients as inputs. The speech corpus consisted of 16 syllables produced by 3 speakers. The input to the neural network was the differences in LPC coefficients sampled at each syllable's time-waveform. A fully connected three-layered back-propagation network was trained by the delta learning rule. With a relatively small number of parameters for each syllable, based on 240 tokens of 16 difficult CV syllables spoken within a sentence context by three speakers, preliminary results for test data indicated that the recognition accuracy is as high as 70.8%.>