CONTINUOUS NUMERAL SPEECH RECOGNITION METHOD

摘要

PURPOSE: To obtain a high recognition rate in recognition of continuous numerals by realizing a context dependent model taking the influence of intonation coupling into consideration while suppressing the required learning data quantity. CONSTITUTION: The acoustic model corresponding to the respective numerals are composed of three partial acoustic models; a model for the part of the head of a word, a model for the part of the middle of a word, and a model for the part of ending of a word at the time of constituting the acoustic model corresponding to the respective numerals by extracting the acoustic characteristic quantity of numeral speeches and statistically modeling the acoustic characteristic quantity thereof. The model for the part of the head of the word and the model for the part of ending of the word among these models are prepd. in plural kinds for the one numeral according to the context of the numerals continuous before and after. The partial acoustic models meeting the context are connected and the characters corresponding to the input speeches are searched from the degree of coincidence with the input speeches at the time of recognition.