METHOD FOR SPEAKER-INDEPENDENT RECOGNITION OF ISOLATED VOICE COMMANDS Russian patent published in 1998 - IPC

FIELD: information processing and control systems, in particular, speech recognition systems. SUBSTANCE: method involves two-level processing of speech signal. First stage of method involves search for possible candidates for signal under analysis while second stage involves selection of most probable choice from found candidates. Novel feature of method consists in analysis of recognition results for first and second stages by means of decision making unit. If given result fails to satisfy criteria of decision making unit, first level processing unit is requested again. First level processing involves search for possible candidates using whole-word analysis method. This involves measuring distance between analyzed speech signal and reference signals using non-linear functions. First level analysis unit measures distance of analyzed speech signal to reference signals from first level unit and threshold level for solving rule in order to select probable candidates as non-linear function from minimal distance to reference signals of first level unit. EFFECT: increased probability of correct recognition, increased speed. 4 cl, 1 dwg