Electrical audio signal processing systems and devices – One-way audio signal program distribution – Public address system
Patent
1989-02-21
1991-11-12
Shaw, Dale M.
Electrical audio signal processing systems and devices
One-way audio signal program distribution
Public address system
381 41, G10L 708, G10L 702
Patent
active
050654314
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to pattern recognition, and particularly, though not exclusively, to speech recognition.
2. Description of the Related Art
The n-tuple method of pattern recognition, which was originally suggested by Bledsoe & Brown ("Pattern recognition and reading by machine", Proc. Eastern Joint Computer Conf., Boston, pp 225-232; 1959), has been proposed for the recognition of two-dimensional patterns. FIG. 1 shows an N.times.M pattern each element of which is represented by a single bit, either a "0" or a "1". Sets of n bits are selected in a specified way (e.g. at random) from the array forming in each case an `n-tuple`. Usually each bit is used once so that there are NM
n-tuples. A template store (FIG. 2, which assumes n=4) has NM
rows (one for each n-tuple) and 2.sup.n columns. In a training sequence each n-tuple is interpreted as a binary number from 0 to 2.sup.n -1 and a 1 is written into the corresponding column of the row assigned to that n-tuple. A number of training passes on patterns, all of course with the same n-tuple selection, will plot further 1's into the template store--which may or may not coincide with those already written in, according to the degree of similarity between the patterns.
Templates are formed in this way for a number of patterns to be recognised. When an unknown pattern is to be identified, n-tuples are formed in the same way and each is used to read out the corresponding location in one of the template stores. The number of `1`s found represents a `score` of the similarity between the unknown and the known pattern. Scores are obtained for each template and the unknown pattern is deemed to be recognised as being that corresponding to the template giving the highest score.
Tattersall and Johnston ("Speech Recognisers based on N-tuple Sampling", Proceedings of the Institute of Acoustics, Vol 6 part 4 pp 405-413, Autumn Conference, 1984) have proposed a speech recogniser using this principle. In this case, in the pattern of figure 1, the column represents successive samples in time of a word of speech, and the bits within each column represent a binary coding (e.g., a bar code) of that sample (or optionally of extracted features). This pattern is then analysed in much the same way as described above.
SUMMARY OF THE INVENTION
This invention stores n-tuple occurrence frequencies in a training mode and uses such stored frequency measurements in a recognition mode to form similarity measurements and further updates the stored frequency measurements during the recognition process in one exemplary embodiment.
The present invention is defined in the appended claims.
BRIEF DESCRIPTION OF THE DRAWINGS
Some embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings in which:
FIGS. 1 and 2 depict prior art pattern recognition processes;
FIG. 3 is a block diagram of one form of speech recognition apparatus according to the invention;
FIG. 4 illustrates the operation of the template store of FIG. 3;
FIGS. 5a and 5b illustrates a modified form of microfeature selection; and
FIG. 6 illustrates a modified version of the template store of FIG. 4.
The embodiment to be described is concerned with speech recognition; however, it is to be noted that the invention is also of value in recognising other types of patterns.
The apparatus shown in FIG. 3 has an audio input 10 for receiving speech signals which are converted by an analogue to digital converter 11. Although the raw time domain samples could be directly utilised (as proposed by Tattersall and Johnston), in this embodiment they are supplied to a front end processing unit 12 which serves to convert them (as described in more detail below) into binary representations of various characteristics of the speech, e.g., energy profile, spectral values, zero-crossing rate. It is assumed that this output is of 80 bits and is provided for each 10 ms period of speech.
The front end processing unit 12 supplies data
REFERENCES:
patent: 3521235 (1970-07-01), Becker
patent: 4087630 (1978-05-01), Browning et al.
patent: 4319085 (1982-03-01), Welch et al.
patent: 4490847 (1984-12-01), Aleksander
patent: 4513436 (1985-04-01), Nose et al.
patent: 4760604 (1988-07-01), Cooper et al.
patent: 4782459 (1988-11-01), Johnston
patent: 4805225 (1989-02-01), Clark
IEEE Transactions on Audio and Electroacoustics, vol. AU-16, No. 2, Jun. 1968, pp. 235-239, IEEE, New York, U.S.; R. F. Purton: "Speech Recognition Using Autocorrelation Analysis".
Steck--"Stochastic Model for the Browing-Bledsoe Pattern Recognition System"--IRE Transactions on Electronic Computers, Apr. 1962, pp. 274-282.
Hughes--"On the Mean Accuracy of Statistical Pattern Recognisors", IEEE Transaction on Information Theory, vol. IT-14, No. 1, Jan. 1968, pp. 55-63.
Ullmann--"Experiments with the N-Tuple Method of Pattern Recognition", IEEE Trans. 1969, C-18, pp. 1135-1137.
Ullmann, Butterworths--"Pattern Recognition Techniques", pp. 112-121.
Aleksander & Stonham--"Guide to Pattern Recognition Using Random Access Memories"--Computers and Digital Techniques, Feb. 1979, vol. 2, No. 1, pp. 29-40.
"Seeking the Mind in Pathways of the Machine"--The Economist, 27 Jun., 1985, pp. 83-88.
Fairhurst & Aleksander--"Dynamics of the Perception of Patterns in Random Learning Nets", MPPP-21, pp. 311-316, date unknown.
Tattersall et al.--"Speech Recognizers Based on N-Tuple Sampling", Proceedings of the Institute of Acoustics, Autumn conference 1984, vol. 6, Part 4, pp. 405-413.
Bledsoe et al., "Pattern Recognition and Reading by Machine"--1959 Proceedings of the Eastern Joint Computer Conference--pp. 225-232.
British Telecommunications public limited company
Merecki John
Shaw Dale M.
LandOfFree
Pattern recognition using stored n-tuple occurence frequencies 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 using stored n-tuple occurence frequencies, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Pattern recognition using stored n-tuple occurence frequencies will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1018414