Improvement of a linear phonetic-context model for phoneme-based continuous speech recognition

A model for coarticulator that considers changes in the degree of coarticularity effects according to the position in a phonetic segment, in a framework of the linear phonetic-context model (LPCM) is described. The position-dependent degree of the coarticulatory effects is represented by a halved Gaussian function, a function of the distance from the phonemic boundary. The standard deviation of the function controls a spread of the coarticulatory effects. The mean-squared error is minimized for a variety of speech data if the standard deviation is set near to 30 or 50 ms. Phoneme recognition experiments using sentences spoken with different speaking rates show that the improved LPCM increases the recognition rates for all cases, and is more robust to the difference of the speaking rate than the original LPCM