Patent
1989-05-09
1992-12-29
Fleming, Michael R.
G01L 914
Patent
active
051757940
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention relates to methods and apparatus for pattern recognition, and particularly, though not exclusively, speech recognition.
The invention is particularly concerned with monitoring temporal sequences of input vectors so as to recognise particular patterns.
2. Description of Related Art
In this specification, an "N dimensional vector" comprises a group of N values, each value corresponding to a respective dimension of the vector. The values may be represented by analogue signals or digitally. Such a vector has a magnitude which may be, for example, defined by the square root of the sum of the squares of the values, and also a direction in N dimensional space. For simplicity, throughout this specification, scalar quantities (except for N) will be represented by lower case letters, and vector quantities by upper case letters.
Vectors of this type can in particular be derived from an analysis of human speech. Thus, an analogue signal representing a series of speech sounds can be regularly sampled and the content of each sample can be represented in terms of a vector comprising a set of feature values corresponding, for example, to the amplitude of respective frequencies within the sample.
A paper entitled "Clustering, Taxonomy, and Topological Maps of Patterns" by T. Kohonen in Proceedings of the Sixth International Conference on Pattern Recognition, October 1982, pages 114-128 describes an approach for the statistical representation of empirical data. Sets (vectors) of input data are successively applied, in parallel to each of a number of processing units regarded as forming a two-dimensional array; each unit produces a single output proportional to the degree of matching between the particular input vector and an internal vector associated with that unit. An adaptation principle is defined so that a succession of input vectors, which form a statistical representation of the input data, cause changes in the internal vectors. This works (for each input vector) by: input (eg the smallest Euclidean distance); neighbourhood; the direction of change being such that the similarity of those internal vectors is increased.
As this `self-organisation` process proceeds, the size of the neighbourhood is progressively reduced; the magnitude of the adjustments may also decrease. At the conclusion of this process, the array internal vectors define a mapping of the input vector space onto the two-dimensional space. Kohonen trained such an array using manually-selected speech samples of certain stationary Finnish vowel phonemes (selected to exclude those including transients), the input vectors each consisting of fifteen spectral values, and found that it mapped the phonemes into the two-dimensional array space.
SUMMARY OF THE INVENTION
According to one aspect of the present invention, there is provided a pattern recognition apparatus comprising:
an input for receiving a temporal sequence of input signal vectors;
store means for storing a plurality of reference vectors;
plurality of comparison elements arranged in operation to receive the same, current, value of the input vector and to generate an output signal indicative of the degree of similarity between the input vector and a respective one of the reference vectors;
means for producing for each element an output modified in dependence on the output signal produced by that element as a result of a comparison between its reference vector and at least the immediately preceding input vector; and
recognition means for comparing the pattern represented by those of the said modified outputs which indicate a relatively high similarity between the input and reference vectors with reference information to identify patterns represented by the said temporal sequence.
This enables the path through the "most similar reference vectors" corresponding to a series of input vectors to be relatively easily determined by building in a certain degree of persistence in the output signals generated for each reference vector. Thu
REFERENCES:
patent: 3287649 (1966-11-01), Rosenblatt
patent: 3845471 (1974-10-01), Reitboeck
patent: 4119946 (1978-10-01), Taylor
patent: 4254474 (1981-03-01), Cooper et al.
patent: 4773024 (1988-09-01), Faggin et al.
patent: 4805225 (1989-02-01), Clark
Transactions of the I.R.E., Professional Group on Information Theory, No. 4, "Simulation of Self-Organizing Systems by Digital Computer" Farley, B. G. and Clark, W. A., Sep. 1954, pp. 76-84.
L'Electricite Electronique Moderne, vol. 42, No. 265, Jun./Jul. 1972, B. H. Marin, "Le neurone cybernetique", pp. 21-25.
Electronics Letters, vol. 4, No. 20, Oct. 4, 1968, I. Aleksander et al. "Microcircuit lerning nets: improved recognition by means of pattern feedback" pp. 425-469.
Pattern Recognition, vol. 15, No. 6, 1982, Pattern Recognition Society, K. Fukushima et al. "Neocognitron: a new algorithm for pattern recognition tolerant of deformations and shifts in position" pp. 455-469.
IEEE Transactions on Computers, vol. 20, No. 9, Sep. 1971, L. C. W. Pols "Real-time recognition of Spoken words" pp. 972-978.
IEEE Journal of Solid-State Circuits, vol. SC-1, No. 2, Dec. 1966, J. W. McConnell et al. "MOS adaptive memory elements as weights in adaptive pattern classifier" pp. 94-99.
IEEE Proceedings of the 6th International Conference on Pattern Recognition, Oct. 1982, T. Kohonen, "Clustering, taxonomy, and topological maps of patterns" p. 18.
IEEE Acoustics, Speech, and Signal Processing Society, Apr. 1987, R. P. Lippmann, "An Introduction to Computing with Neural Nets" pp. 4-22.
Kybernetik, vol. 16, No. 2, 1974, pp. 103-112; H. Wigstrom: "A model of a neutral network with recurrent inhibition".
Systems, Computers, Controls, vol. 6, No. 5, 1975, pp. 15-22; K. Fukushima: "Self-organizing multilayered neural network".
British Telecom Technology Journal, vol. 6, No. 2, Apr. 1988, G. D. Tattersall et al: "Neural arrays for speech recognition", pp. 140-163.
British Telecommunications public limited company
Doerrler Michelle
Fleming Michael R.
LandOfFree
Pattern recognition of temporally sequenced signal vectors does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Pattern recognition of temporally sequenced signal vectors, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Pattern recognition of temporally sequenced signal vectors will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1893243