Patent
1994-03-18
1996-09-10
Downs, Robert W.
395 24, 395 264, 395 265, G10L 506, G10L 900
Patent
active
055553443
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
The problem arises in various fields of pattern recognition of having to make use of multidimensional feature vectors whose individual components, the features, are relevant in different ways in the case of different patterns to be recognized. This situation occurs, in particular, in automatic speech recognition, in which when previous recognition systems are used it is easy for phonetically similar words, (for example, German words "zwei" and "drei") to be confused. It is particularly easy when using known recognition systems to confuse words which differ only in a single phoneme (for example, German phonemes "dem" and "den"). This problem becomes still more acute in the case of speaker-independent recognition of speech which is carried over telephone lines, because due to the reduced transmission bandwidth of 3.4 kHz speech-relevant frequency ranges are lost (for example, The sounds /s/ and /f/ can no longer be distinguished over the telephone).
Some of these known recognition systems are based on a direct pattern comparison of stored reference words and the actually spoken word, with account being taken of temporal fluctuations in rate of speech. These fluctuations are taken into account with the aid of dynamic programming, for example. Moore has proposed an approach for such recognition systems (R. K. Moore, M. J. Russel, M. J. Tomlinson, "The discriminative network: A mechanism for focusing recognition in whole word pattern matching", IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 1041-1044, Boston, 1983, ICASSP), which automatically finds the discrimination-relevant parts of words and weights these more strongly by comparison with the other parts. A disadvantage of this method is that the automatic search of discrimination-relevant parts can be affected by errors in the case of confusable word pairs. Discrimination-relevant word parts are not always found, or word parts are wrongly regarded as discrimination-relevant. This problem also cannot be solved in principle using the method of dynamic programming alone.
SUMMARY OF THE INVENTION
In other fields of signal processing and pattern recognition, very similar problems occur as in the field of speech recognition. It is therefore the object of the invention to specify a method for recognizing patterns in time-variant measurement signals, by means of which the frequency of the confusion of similar feature vectors can be substantially reduced. This object is achieved with the aid of a method for recognizing patterns in time-variant measurement signals by classifying a temporal sequence of pattern vectors and by reclassification in pairs.
In this method, the sequence of feature vectors which is to be classified is segmented with the aid of a Viterbi decoding algorithm by comparing this sequence to be classified with a set of hidden Markov models. For this purpose, there is calculated for each hidden Markov model a total emission probability for the generation of the sequence to be classified by this hidden Markov model. Subsequently, an optimum assignment path from feature vectors to states of the hidden Markov models is determined by backtracking.
A discriminating comparison of the assignments is carried out for selected or all pairs of hidden Markov models by calculating modified total emission probabilities for each hidden Markov model of a pair on the assumption that the respective other hidden Markov model of the same pair competes with the hidden Markov model under review, and by determining the respective more probable hidden Markov model of a pair. Thereafter, the hidden Markov model with the largest total emission probability is selected from among all the pairs under review.
The method has the advantage that a pattern to be classified is compared not with a reference pattern but with a statistical distribution function of many reference patterns. In this way, it is not a simple distance between two patterns to be recognized which is obtained, as is the case with dynamic programming,
REFERENCES:
patent: 4677672 (1987-06-01), Ukita et al.
patent: 4741036 (1988-04-01), Bahl et al.
patent: 4803729 (1989-02-01), Baker
patent: 4975959 (1990-12-01), Benbassat
patent: 5170432 (1992-12-01), Hackbarth et al.
patent: 5228087 (1993-07-01), Bickerton
patent: 5333275 (1994-07-01), Wheatley et al.
patent: 5390278 (1995-02-01), Gupta et al.
patent: 5425129 (1995-06-01), Garman et al.
"An Introduction to Hidden Markov Models", L. R. Rabiner, B. H. Juang, IEEE Transactions on Acoustics, Speech and Signal Processing, Jan. 1986, pp. 4-16.
"Mathematical Methods of Feature Selection in Pattern Recognition", J. Kittler, International Journal, Man-Machine Studies, (1975), pp. 609-637.
Nakaga wa et al, "A method for continuous speech segmentation using HMM"; 9th Internation Conference on Pattern Recognition, pp. 960-962, 1988.
Ma et al, "TDNN Labeling for a HMM Recognizer"; ICASSP '90, pp. 421-423, 1990.
Brummer et al, "Automatic speaker independent alignment of continuous speech with its phonetic transcription using a hidden markov model"; COMSIG 88, pp. 35-40, 1988.
Downs Robert W.
Hafiz Tariq
Siemens Aktiengesellschaft
LandOfFree
Method for recognizing patterns in time-variant measurement sign 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 for recognizing patterns in time-variant measurement sign, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for recognizing patterns in time-variant measurement sign will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1327214