Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Reexamination Certificate
2000-05-31
2002-08-20
Dorvil, Richemond (Department: 2654)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
C704S231000
Reexamination Certificate
active
06438519
ABSTRACT:
FIELD OF THE INVENTION
This invention relates to pattern classification, such as speech recognition in trained systems.
More particularly, the present invention relates to apparatus and method for the rejection of out-of-class inputs.
BACKGROUND OF THE INVENTION
Pattern classification systems can be used in two typical situations. A closed set situation involves a fixed set of known classes. Given an input, the classifier is required to pick the best choice from this list of classes. Alternatively, a classifier may be used in an open set manner. A general scenario for open set classification is that the classifier is presented with feature vectors from a single class. The classifier then determines if the features are from a previously known class or an unknown class. An application of the open set problem is out-of-vocabulary rejection in speech recognition. Speech recognition is used as an example herein because of its widespread knowledge and use, but it will be recognized by those skilled in the art that virtually all pattern recognition systems are equally applicable. In the speech recognition case, the recognizer has a known vocabulary; typically, the user would prefer flagging of an unknown word to misrecognition.
Using a speech recognition system as a typical example, assume that in a specific instance the system is looking for a “yes” or a “no”. In many different situations, the user might utter some out-of-vocabulary sounds, such as “oh”, “ah”, “er”, or the user might cough or clear his throat. Typically, the speech recognition system looks at components of the utterance, compares them to components of the words it is looking for, i.e., yes and no, and uses a threshold to determine whether the utterance is sufficiently close to one of the words to be positively recognized. Here the problem is two-fold. In many instances the utterance may be so close,(e.g. “oh” and “no”) it is misclassified. To add to this problem, in a noise system or under noise conditions, much of an utterance may be masked or lost. Thus, while the threshold remains constant, the masking (e.g. noise, closeness of the received word, or signal, etc.) may vary substantially under different operating conditions.
Current speech recognition applications experience significant limitation in customer acceptance due to inadequate rejection of out-of-class input. However, reject options for decision rules are not a new area for pattern recognition. The optimum rejection rule for pattern recognition (Bayes rejection rule) was introduced over 30 years ago in an article by C. K. Chow, entitled “On optimum recognition error and reject tradeoff”,
IEEE Trans. Inf. Theory
”, IT-16, no. 1, pp. 41-46, January 1970. Nevertheless, this work, and extension to it, assume we have exact knowledge of the class statistics.
In the case of speech recognition, only an estimate of the class statistics is available to the pattern recognition system. Furthermore, when the input speech is corrupted by noise, resulting in mismatched conditions, the original probability distribution estimates are no longer good approximations to the actual distributions. Thus, the optimal Bayes reject rule in quiet conditions is no longer valid.
Recent work to improve the rejection criterion for out-of-vocabulary words has focused on likelihood ratios between in-class model scores, as well as garbage, or filler, models to model the out-of-class feature space. See for example: C. S. Ramalingam et al., “Speaker-dependent name dialing in a car environment with out-of-vocabulary rejection”,
Proc. ICASSP,
pp. 165-168, 1999; A. Bayya, “Rejection in speech recognition systems with limited training”,
Proc. ICSLP,
1998; H. Boulard, B. D'hoore and J. M. Boite, “Optimizing recognition and rejection performance in wordspotting systems”,
Proc. ICASSP,
pp. I-373-I-376, 1994; and R. C. Rose and D. B. Paul, “A hidden Markov model based keyword recognition system”,
Proc. ICASSP.
pp. 129-132, 1990. However, a score based threshold is still employed in order to provide a mechanism for trading-off recognition and rejection error rates based on some cost function. The use of a threshold leads to significant performance degradation in severely mismatched conditions. This is due to the altered statistics of the input features, resulting in a compression of the score range.
Accordingly it is highly desirable to provide pattern recognition apparatus and a method of overcoming these problems. Further, it is highly desirable to provide apparatus and methods of overcoming the problems without substantially increasing the computations required or the amount of memory used.
REFERENCES:
patent: 5233681 (1993-08-01), Bahl et al.
patent: 5428707 (1995-06-01), Gould et al.
patent: 5572624 (1996-11-01), Sejnoha
patent: 5680511 (1997-10-01), Baker et al.
patent: 5729656 (1998-03-01), Nahamoo et al.
patent: 5737489 (1998-04-01), Chou et al.
patent: 5805772 (1998-09-01), Chou et al.
patent: 5897616 (1999-04-01), Kanevsky et al.
patent: 5937383 (1999-08-01), Ittycheriah et al.
patent: 5946653 (1999-08-01), Campbell et al.
patent: 5950159 (1999-09-01), Knill
patent: 5963940 (1999-10-01), Liddy et al.
patent: 6006221 (1999-12-01), Liddy et al.
patent: 6161090 (2000-12-01), Kanevsky et al.
patent: 6188976 (2001-02-01), Ramaswamy et al.
patent: 6192360 (2001-02-01), Dumais et al.
patent: 6195634 (2001-02-01), Dudemaine et al.
patent: 6263308 (2001-07-01), Heckerman et al.
Broun Charles Conway
Campbell William Michael
Dorvil Richemond
Koch William E.
Motorola Inc.
Opsasnick Michael N.
LandOfFree
Apparatus and method for rejecting out-of-class inputs for... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Apparatus and method for rejecting out-of-class inputs for..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Apparatus and method for rejecting out-of-class inputs for... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2947536