A FEATURE-BASED HIERARCHICAL SPEECH RECOGNITION SYSTEM FOR HINDI

This paper presents a description of a speech recognition system for H indi. The system follows a hierarchic approach to speech recognition a nd integrates multiple knowledge sources within statistical pattern re cognition paradigms at various stages of signal decoding. Rather than make hard decisions at the level of each processing unit, relative con fidence scores of individual units are propagated to higher levels. Ph oneme recognition is achieved in two stages: broad acoustic classifica tion of a frame is followed by fine acoustic classification. A semi-Ma rkov model processes the frame level outputs of a broad acoustic maxim um likelihood classifier to yield a sequence of segments with broad ac oustic labels. The phonemic identities of selected classes of segments are decoded by class-dependent neural nets which are trained with cla ss-specific feature vectors as input. Lexical access is achieved by st ring matching using a dynamic programming technique. A novel language processor disambiguates between multiple choices given by the acoustic recognizer to recognize the spoken sentence.