Method and neural network for speech recognition using a correlo

Details Method and neural network for speech recognition using a correlo Method and neural network for speech recognition using a correlo

1994-01-21

1998-02-24

MacDonald, Allen R.

395 268, G10L 506

Patent

active

057218072

DESCRIPTION:

BRIEF SUMMARY
BACKGROUND OF THE INVENTION



Field of the Invention

The invention relates to a method for recognizing individual words of spoken language, and an apparatus for performing the method.
The recognition of spoken speech is among the most important fields in communications technology for the future, yet at the same time is among the most difficult. A great number of applications for speech recognition exist, such as spoken input rather than keyboard input in text processing systems (phonetic typewriters), information retrieval or ordering over the telephone, and speech control of machine systems. Thus far, however, the widespread introduction of speech recognition systems has failed because of a number of unsolved problems that are both technical and psychological in nature.
Speech recognition is a task of pattern recognition. The speech signal of the spoken expression (sound-word-sentence, etc.) forms the input pattern, whose significance must be recognized or understood.
The recognition process can be broken down into two stages: feature extraction and classification. Feature extraction serves to extract the characteristics that are relevant to recognition from the speech signal and to eliminate portions that are not relevant. In the second step classification conclusions--as to the significance of the present speech sample are drawn from the present extracted characteristics.
The object of the invention is to disclose a method with which a limited number of individual words of spoken speech can be recognized.


SUMMARY OF THE INVENTION

This is done in accordance with the invention by a method defined by claim 1.
The method of the invention is distinguished by especially high recognition rates. Moreover, it is relatively insensitive to background noise. A further advantage of the method is that the correlogram K on which the classification of the individual words is based can be defined simply. This simple calculatability is a prerequisite for achieving a speech recognition system at little expense for hardware. Another advantage of the method is that the number of recognizable individual words is not fundamentally limited. The method of the invention is therefore especially advantageous in applications involving a very large vocabulary.
A favorable compromise between the amount of data to be processed--the expenditure for calculation--and a high recognition rate is achieved by embodying the method as defined by the body of claim 2.
Especially high recognition rates are achieved if the conditions as defined by the body of claim 3 are chosen for the indices j, h, k of the correlogram K.
The use of a neural network for classifying the spoken individual word has further advantages. Neural networks are a rough simulation of the human brain structure, with its associative mode of functioning. In problems of pattern recognition, of the kind that also exist in speech recognition, they are superior to conventional computer structures.
The neural network as defined by the body of claim 5 is distinguished by its especially good "learnability"; that is, the "training phases" for the words to be recognized are reduced to a minimum.
A feature of the method as defined by claim 6 brings about simple attainability of the neural network and a further reduction in the expenditure for calculation.
With the apparatus of claim 7, the method according to the invention can be attained especially favorably.
For use in higher-order systems, it is favorable if the entire apparatus defined by claim 8 is embodied as an integrated component.
The use of a speech recognition method in a telephone set in accordance with claim 9 not only brings about great convenience in use but also, when used in a car phone, an increase in traffic safety, since the driver and telephone user is not diverted by the dialing process.
The invention will be described in further detail in conjunction with four figures.


BRIEF DESCRIPTION OF THE DRAWING

FIG. 1, a course of the method of the invention;
FIG. 2, the spectral amplitude distribution of three diff

REFERENCES:
patent: 4715065 (1987-12-01), Parker
patent: 4975961 (1990-12-01), Sakoe
patent: 5040215 (1991-08-01), Amano et al.
patent: 5285522 (1994-02-01), Mueller et al.
patent: 5404422 (1995-04-01), Sakamoto et al.
patent: 5426745 (1995-06-01), Baji et al.
patent: 5473759 (1995-12-01), Slaney et al.
Patent Abstracts of Japan, vol. 10, No. 340, Nov. 18, 1986, JP 61-144157, Jul. 1, 1986.
Rumelhart et al., "Parrallel Distributed Processing", p. 328, 1986 Massachusetts Institute of Technology.
Behme, "A Neural Net for Recognition and Storing of Spoken Words", Parrallel Processing in Neural Systems and Computers, Elsevier Sci. Pub. 1990.
Kowalewski et al., "Word Recognition with a Recurrent Neural Network", Parrallel Processing in Neural Systems and Computers, Elsevier Sci. Pub. 1990.
Komori et al., "Combining Phoneme Identification Neural Networks into an Expert System Using Spectrogram Reading Knowledge", ICCASSP '90, IEEE Acoustics, Speech and Signal Processing Conference, 1989.
Hatazaki et al., "Phoneme Segmentation Using Spectrogram Reading Knowlege", ICCASSP '89, IEEE Acoustics, Speech and Signal Processing Conference, 1989.
Komori et al., "Robustness of a Feature Based Phoneme Segmentation System to Speaker Independent and Countinuous Speech", ICCASSP '91, IEEE Acoustics, Speech and Signal Processing Conference, 1991.
Palakal et al., "Feature Extraction from Speech Spectrograms Using Multi-Layered Network Models", Tools for Artificial Intelligence, 1989 Int'l Workshop.
Patent Abstracts of Japan, vol.10, No.340, Nov. 18, 1986 JP 61-144157 -JP 61-144159, Jul. 1, 1986.
Patent Abstracts of Japan, vol.14, No.264, Jun. 7, 1990 JP 2-72396 -JP 2-72398, Mar. 12, 1990.
Muthusamy et al, "Speaker-Independent Vowel Recognition: Spectrograms Versus Cochleagrams", ICASSP '90: Acoustics, Speech and Signal Processing Congerence 1990.

Affiliated with

Also associated with

Rating

Method and neural network for speech recognition using a correlo does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Method and neural network for speech recognition using a correlo, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and neural network for speech recognition using a correlo will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-1879970