SPEECH RECOGNITION DEVICE AND METHOD USING WEIGHING VALUE PER STATE

摘要

PURPOSE: A speech recognition device and a method using weighing values per state are provided to be capable of remarkably reducing error recognition rate of speech by applying a context dependent state weighing value. CONSTITUTION: A characteristic vector extractor(20) receives speech signals of a speaker, extracts the characteristic vector of the speech signal at a fixed frame rate, and adds same weighing values to the characteristic vectors pre the extracted frame. A frame information rate extractor(30) applies weighing value per hidden makov states to respective frames through a multi layer percept theory learned by a divisional method from a learning database, and generates a characteristic vector with a context dependent information rate for respective frames. A speech recognizer(22) recognizes the characteristic vector with a context dependent information rate using a hidden makov model and a learning model.