An isolated word recognition system based on acoustic-phonetic analysis and statistical pattern recognition

A polynomial discriminant function is used to establish the probability density function for voice/unvoice/silence parts of speech. Based on these densities, segmentation accuracy of 95% were obtained. Voice segments are further segmented into phonemic units using threshold functions based on energy and first formant changes (80% accuracy). Multi-dimensional probability density functions based on LPC, energy, and zero crossing serves as prototype for each phonemic unit. Prototypes are also establish for a set of phoneme-pairs. Bayes' rule is used to assign probabilities for each phoneme and phoneme-pair in the unknown speech. Word Recognition is achieved by finding the word with the highest score for its phonemic units.