Data processing: artificial intelligence – Neural network – Structure
Reexamination Certificate
2001-03-01
2004-09-28
Davis, George B. (Department: 2121)
Data processing: artificial intelligence
Neural network
Structure
C706S030000, C706S020000
Reexamination Certificate
active
06799171
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to a system of speech recognition. More specifically, this invention relates to a logic unit and a system of speech recognition with the aid of a neural network. This invention also relates, however, to a new neural network for applications other than speech recognition.
2. Discussion of the Background
Methods of performing speech recognition are of crucial importance in particular for the development of new telecommunications services. The qualities required of a speech recognition system are in particular the following:
Precision—systems making it possible to recognize correctly less than a very high percentage, for example less than 85 percent, of words have only few practical applications.
Insensitivity to noise—the systems must allow a satisfactory recognition even in a noisy environment, for example when the communications are transmitted through a mobile telephone network.
Large vocabulary—for a lot of applications it is necessary to be able to recognize a high number of different words—for example more than 5000.
Independence of speaker—a lot of applications require satisfactory recognition regardless of who the speaker is, and the same for speakers unknown to the system.
The known systems of speech recognition generally carry out two distinct tasks. A first task consists in converting the voice into a digital signal and of extracting a sequence of vectors of voice parameters from this digital signal. Different systems are known for executing this task which generally allow conversion of each frame, of 10 milliseconds for example, of voice into a vector (“features vector”) containing a group of parameters describing at best this frame in the time and frequency domain.
The second task consists in classifying the sequence of vectors received by means of a classifier and establishing to which class (corresponding, for example, to phonological elements such as phonemes, words or sentences, for example) they correspond with the greatest probability among the classes defined during a learning phase of the system. The problem for classifiers is thus to determine, for each input speech vector, the probability of belonging to each defined class.
The speech recognition systems most widely used at the present time use a classifier functioning with the aid of hidden Markov models, better known by the Anglo-Saxon designation Hidden Markov Models (HMM), and illustrated by
FIG. 1
a.
This statistical method describes the voice through a sequence of Markov states
8
1
,
8
2
,
8
3
. The different states are connected by links
9
1
-
9
6
indicating the probabilities of transition from one state to another. Each state emits a voice vector, with a given probability distribution. A sequence of states, defined a priori, represents a predefined phonological unit, for example a phoneme or a triphone. A description of this method is given, for example, by Steve Young in an article entitled “A Review of Large-Vocabulary Continuous-Speech Recognition,” published in September 1996 in the
IEEE Signal Processing Magazine.
In spite of a very poor modelling of time relations between successive speech vectors, this method currently offers the best rates of recognition.
Other systems of classification, which have made it possible to achieve a certain success, use networks of artificial neurons, such as illustrated in
FIG. 1
b,
in particular time delay neural networks (TDNN—Time Delay Neural Networks) or recurrent neural networks (RNN—Recurrent Neural Network). Examples of such systems are described in particular by J. Ghosh et al. in “Classification of Spatiotemporal Patterns with Applications to Recognition of Sonar Sequences” in
Neural Representation of Temporal Patterns,
pages 227 to 249, edited by E. Covey et al., Plenum Press, New York, 1995. All these systems use a delay line comprising registers
25
for the input speech vectors
2
as well as delay elements
26
in their architecture. Computing elements
11
(neurons), interconnected (by means of synapses) with the registers
25
and organized in a hierarchical manner, allow particular phonological elements to be identified. These systems also make it possible to model the time relation between past information and current information, and to correct certain weaknesses of HMMs, without, however, succeeding in replacing them completely.
A more recent approach consists in combining the HMMs with neural networks in hybrid speech recognition systems. Such systems are described, for example, by H. Boulard et al. in “Connectionist Speech Recognition—A Hybrid Approach,” 1994, Kluwer Academic Publishers (NL). These systems have the advantage of a better modelling of context and of phonemes than the HMMs. The price to pay for these systems, however, is either a long training time, due to the error back propagation (EBP) algorithm used, or a limited number of weighting coefficients available for modelling the speech signals.
Another approach is disclosed in the American U.S. Pat. No. 5,220,640 of Jun. 15, 1993. This document describes a neural network architecture by which the input signal has been scaled differently by a “time-scaling network.” The output signals indicate how the entering signals have been changed in scale correspond to learned patterns.
These different systems generally model each word as a sequence of phones, and are optimized to identify each phone in a speech signal as precisely as possible. A correct identification of each phone ensures in principle a perfect recognition of words or of sentences—insofar as these words and these sentences are correctly modelled. In practice, all these systems have the drawback of a lack of robustness in noisy conditions or of results of variable quality, as indicated in particular by S. Greenberg in “On the origins of speech intelligibility in the real world,” ESCA-NATO Tutorial and Research Workshop on Robust Speech Recognition for Unknown Communication Channels, 17
th
-18
th
Apr. 1997, Pont-à-Mousson, France, and by Steve Young in the article indicated further above.
BRIEF SUMMARY OF THE INVENTION
One object of the present invention is thus to propose a system and a method of speech recognition that avoids the drawbacks of prior art systems and methods. More specifically, an object of the present invention is to propose a classifier and a method of classifying speech vectors, improved over prior art classifiers and classification methods.
Another object of the present invention is to improve the performance of a classifier without adding substantially to the complexity, in particular without adding substantially to the number of computing elements.
According to the invention, these various objects are achieved thanks to the features of the independent claims, preferred variants being indicated moreover in the dependent claims.
The invention begins with the observation that speech is more than a linear succession of phones of equal importance for recognition. Experience has shown that even experienced listeners struggle to identify more than 60% of phones presented in isolation; only the context permits the human brain to comprehend sentences and to identify, a posteriori, each phone.
The invention puts this discovery to use by suggesting, for the recognition of speech, integration of features of speech segments much longer than that done in the prior art—for example features of several syllables, of a whole word, even of several words or even of an entire sentence.
To avoid adding to the complexity of the system and the number of computing elements, a hierarchical architecture is proposed, with a system of several tiers. Each tier comprises at least one spatiotemporal neural network (STNN). The rate of signals input in the different tiers of the system is variable, in such a manner that the rate of speech vectors input in the lower tier is adapted, for example, to the recognition of isolated phones, or other brief phonological elements, whereas the rate of signals applied on the upper tiers permits, for exampl
Davis George B.
Oblon & Spivak, McClelland, Maier & Neustadt P.C.
Swisscom AG
LandOfFree
Applicator and method for combating pests, especially... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Applicator and method for combating pests, especially..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Applicator and method for combating pests, especially... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3242112