Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Reexamination Certificate
1999-06-22
2002-04-16
Dorvil, Richemond (Department: 2641)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
Reexamination Certificate
active
06374221
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to advantageous aspects of an improved speech recognizer. More particularly, such a speech recognizer can be used for voice operated name or digit dialing in a telephone device such as a cellular or cordless phone, or also for voice controlled operation of the telephone device generally or any other suitable voice controlled device. The speech recognizer can be trained for speaker independent or speaker dependent operation. In case of a telephony system, the speech recognizer can also be incorporated at the network side of the system, in the network itself or at a particular service providing node of the network.
BACKGROUND OF THE INVENTION
The article “A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition”, Laurence R. Rabiner, Proceedings of the IEEE, Vol. 77, No. 2, February 1989, pp. 257-286 discloses basic principles of speech recognition, and general implementations of speech recognizers for recognition of continuous speech, or for recognition of isolated words or clusters of connected words. This article is incorporated by reference herein in its entirety. Speech recognition of continuous speech, typically involving a large vocabulary of words, is a complicated process, whereas recognition of isolated words, used for small vocabulary word recognition in voice driven control of apparatuses, or, in telephony applications for name or digital dialing and/or control, for instance, is a much simpler process to be implemented.
In said article, methods are disclosed for modeling speech signals to be recognized into statistical models, and, after training of these models, for recognizing unknown speech signals by matching such unknown signals to the models and by taking a statistically most likely model as a recognition result. The recognition result can be used for many applications, including voice command driven operation of apparatuses and, in telephones, name dialing. As described in said article, the speech recognition process is based upon statistically matching feature vectors, representing samples of a speech signal to be recognized, with speech signal models. Such models are statistical models of speech signals, i.e., are models which characterize the statistical properties of the speech signals, e.g., in the form of so-called Hidden Markov Models, HMMs, as described in detail in said article.
Hidden Markov Models are probabilistic functions of so-called Markov chains of states, representing a real-world speech signal as a parametrical stochastic process, the parameters initially being obtained by training of the model using a succession of the same known utterance, for instance. In such a HMM, used in an isolated word speech recognizer, for instance, an observation representing a speech signal is a probabilistic function of the states of the Markov chain, i.e., the resulting model is a doubly embedded stochastic process with an underlying stochastic process that is not observable (hidden), but can only be observed through another set of stochastic processes that produce the sequence of observations. Observations can represent spectral and/or temporal characteristics of a speech signal. A spectral method for obtaining an observation sequence or feature vector sequence from speech samples is so-called LPC/Cepstral feature analysis, as described on page 227 of said article. Typically, a feature vector comprises some 25 different characteristics characterizing the speech signal to be recognized. In the speech recognizer, given an observation sequence, or vector, derived from an unknown input utterance, and the models, probabilities are computed for such an observation sequence as regards all models, i.e., scores are determined. The model with the best score, in terms of likelihood, is selected as a tentative recognition result, which can either be accepted or rejected. While determining scores, using a Viterbi algorithm, model state transitions are saved for later use in a back tracking process for determining a corresponding state sequence best explaining the observation sequence. For retraining or re-estimating the model, the observation sequence and the saved model state transitions are used. The models can be retrained during normal operation of a speech recognizer, i.e., based upon speech signals to be recognized and recognition transcripts. Thus, the models can be improved in a statistical sense so as to improve speech recognition. A transcript, sometimes called a label is the verbal content of an utterance, or an index designation of such a content.
In the AT&T Owner's Manual “AT&T 3050/3450 Cellular Telephone”, pages 59-65, published 1993, Voice Recognition Features of a Voice Dialer are described, such as training the voice dialer before use, of both voice commands for operating the voice dialer and for constructing a voice dial name list.
In the PCT Patent Application WO 98/25393, a voice dialer is disclosed. A telephone directory of names and telephone numbers which a user frequently dials is maintained. A speech recognizer is used to select records from the directory. The speech recognizer performs similarity measurements by comparing confidence metrics derived from an input representation of a spoken utterance and from a previous representation stored in a record of the directory. As described on page 22, lines 5-14, the user's own dialing habits are used to improve selection of the desired record, and to organize and update a telephone directory. In this respect, the user's calling habits are registered by recording in a frequency field the number of times the user calls a particular number. Furthermore, adding and deleting of records is controlled by frequently prompting the user, to speak again or to cancel, for instance, so as to avoid possible errors in the telephone directory.
In other speech recognition systems, the reliability of the tentative recognition result, obtained by selecting the statistically most likely model in the matching process, is tested. If in a number of tests most or all tests are passed, the tentative recognition result, which is rejected otherwise, is accepted as the final recognition result. One known test is the testing of a so-called anti-model, an anti-model being a model which represents a recognition outcome which can easily be statistically confused with a model of interest. If in such a test the anti-model does not score significantly worse than the model of interest, a user input utterance under test is rejected. In some other systems, this step is optional, and the tentative recognition result becomes the final recognition result without any reliability testing.
Furthermore, as regards training of speech recognition models, it is well-known that in order to get good performance of a set of models, both for sets trained with utterances of many speakers, so-called speaker independent training, and for sets trained with utterances of an individual user, so-called speaker dependent training, the training set should be sufficiently large and preferably should include the conditions under which the models are to be used during normal operation of the speech recognizer, such conditions including ambient noise, transducer and channel distortion, the user's emotional state, and so on. In speaker independent training, the training set should include utterances from speakers having various pronunciations of a particular language. In practice, it is difficult to fulfill all of the above requirements. Collecting training sets for speaker independent training is expensive, tedious and time consuming. Collecting more than some two utterances per word for speaker dependent recognition is considered by many users as an imposition. Furthermore, the pronunciation of a single user may vary with time. Moreover, it is difficult to foresee all the different noise and channel conditions in which the speech recognizer will be used in practice, and to include such conditions in the training set. In addition, the emotional state of people traini
Dorvil Richemond
Lucent Technologies - Inc.
LandOfFree
Automatic retraining of a speech recognizer while using... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Automatic retraining of a speech recognizer while using..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Automatic retraining of a speech recognizer while using... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2919229