Speech recognition system and program thereof

A speech recognition apparatus comprising:

• a characteristic quantity extraction unit for extracting a characteristic quantity of an inputted speech to be recognized, said apparatus for performing speech recognition by matching between a predetermined speech and a phoneme hidden Markov model of speech data previously recorded;
• a composite model generation unit for generating a composite model by synthesizing the phoneme hidden Markov model of the speech data and a hidden Markov model of noise data previously recorded;
• and
• a speech recognition unit for recognizing the inputted speech by matching the characteristic quantity being extracted in the characteristic quantity extraction unit from the inputted speech, to the composite model generated in the composite model generation unit, wherein the speech recognition unit executes matching between the characteristic quantity and the composite model to be matched thereto while dynamically selecting the composite model in response to changes of noise associated with the inputted speech.

A speech recognition apparatus comprising:

• a speech database storing speech data as models for speech recognition;
• a noise database storing noise data assumed to generate under a predetermined noise environment;
• a composite model generation unit for generating a composite model by synthesizing a speech model generated based on the speech data read out from the speech database and a noise model generated based on the noise data read out from the noise database;
• and
• a speech recognition unit for performing speech recognition by matching between a characteristic quantity of an inputted speech to be recognized and the composite model generated in the composite model generation unit independently of each speech frame of the inputted speech.

A speech recognition apparatus comprising:

• a speech database storing speech data as models for speech recognition;
• a noise database storing noise data assumed to generate under a predetermined noise environment;
• a composite model generation unit for generating a composite model by synthesizing a speech model generated based on the speech data read out from the speech database and a noise model generated based on the noise data read out from the noise database;
• and
• a speech recognition unit for performing speech recognition by matching between a characteristic quantity of an inputted speech to be recognized and the composite model generated in the composite model generation unit while dynamically selecting the composite model to be matched thereto in response to changes of noise associated with the inputted speech in speech.

A speech recognition method for recognizing a speech by controlling a computer, the method comprising the steps of:

• extracting a characteristic quantity of an inputted speech to be recognized and storing the characteristic quantity into a memory;
• reading out from a memory a phoneme hidden Markov model into which noise elements are incorporated, the noise elements being generated based on noise data and predetermined speech;
• and
• recognizing the inputted speech by matching the characteristic quantity of the inputted speech stored in the memory to the phoneme hidden Markov model for each speech frame of the inputted speech, while the recognition being performed based on results of the matching.

A speech recognition method for recognizing a speech by controlling a computer, the method comprising the steps of:

• extracting a characteristic quantity of an inputted speech to be recognized and storing the characteristic quantity into a memory;
• reading out from a memory a phoneme hidden Markov model into which noise elements being generated based on noise data and predetermined speech are incorporated;
• and
• recognizing the inputted speech by matching the characteristic quantity of the inputted speech stored in the memory to the phoneme hidden Markov model while dynamically selecting the phoneme hidden Markov model to be matched thereto in response to changes of the noise associated with the inputted speech.

A computer readable storage medium in which a program for controlling a computer to execute speech recognition is recorded, the program making the computer function as:

• characteristic quantity extraction means for extracting a characteristic quantity of an inputted speech to be recognized;
• composite model generation means for generating a composite model by synthesizing a phoneme hidden Markov model of speech data previously recorded and a hidden Markov model of noise data previously recorded;
• and
• speech recognition means for recognizing the inputted speech by matching between the characteristic quantity of the inputted speech and the composite model for each speech frame in the inputted speech.

A computer readable storage medium in which a program for controlling a computer to execute speech recognition is recorded, the program making the computer function as:

• characteristic quantity extraction means for extracting a characteristic quantity of an inputted speech to be recognized;
• composite model generation means for generating a composite model by synthesizing a phoneme hidden Markov model of speech data previously recorded and a hidden Markov model of noise data previously recorded;
• and
• speech recognition means for recognizing the inputted speech by matching the characteristic quantity of the inputted speech to the composite model while dynamically selecting the composite model to be matched thereto in response to changes of the noise associated with the inputted speech.