Signal processing arrangements

• US 5,519,805 A
• Filed: 04/18/1994
• Issued: 05/21/1996
• Est. Priority Date: 02/18/1991
• Status: Expired due to Term

• Current Assignee: Hydralogica IP Limited, John Jenkins
• Original Assignee: Domain Dynamics (Holdings) Ltd.
• Inventors: King, Reginald A.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Doerrler, Michelle
• Application Number:
  US08/229,062
• Time in Patent Office: 764 Days
• Field of Search: 395/2.4-2.66, 395/21-22, 395/2-2.11, 381/45-53
• US Class Current: 704/202
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 17/18 Artificial neural networks; Connectionist approaches
• CPC Class Codes: G10L 19/00 Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis in musical instruments G10H

Abstract

A signal processing system for time varying band-limited signals such as speech comprises signal coding means for affording a time encoded signal symbol stream, a matrix generator for generating a fixed size, typically "A" matrix from the symbol stream, and an artificial neural network to which the matrix elements are applied for network training and subsequently for affording an output indicative of the nature of an input signal.

Claims

1. A signal processing system for discriminating time varying band-limited input signals comprising means for generating a plurality of time varying band-limited input electrical analog signals such as speech from a plurality of entities to be discriminated, a signal coder for time encoding each of the input signals into a plurality of time encoded symbols to afford a time encoded signal symbol stream, a matrix generator for generating a fixed size matrix composed of a plural coordinate array of cells, said matrix generator receiving each of the time encoded signal symbol streams and coding each said signal symbol stream into the cells of a fixed size matrix generated by the matrix generator, and said matrix generator having a plurality of outputs correlated in number with the cells of the fixed size matrix,

a matrix data store for storing multiple sample matrices generated by the matrix generator, and

an artificial neural network processor comprised of an input layer, a hidden layer, an output layer,

the input layer of the ANN (artificial neural network) having a plurality of inputs correlated in number with the plurality of outputs from the matrix data store for receiving the contents of individual cells of the fixed size matrix,

the output layer of the ANN having a plurality of outputs correlated to the entities to be discriminated,

the hidden layer of the ANN having a plurality of inputs and outputs correlated to and connected to the inputs of the input layer and the outputs of the output layer, respectively,

the multiple sample matrices stored in the matrix data store being indicative of each of the plurality of entities from which the characteristics of the ANN can be set to effect discrimination as between the entities.

2. A signal processing system according to claim 1 further comprising a skeletonizing means for employing a parameter of the cell contents of the fixed size matrices to skeletonize the outputs from the matrix generator to substantially reduce the requirements for inputs to the input layer of the ANN.

3. A signal processing system according to claim 2 wherein the parameter is a preselected value for cell content whereupon the contents of all cells are discarded which are less than said preselected value.

4. A signal processing system according to claim 1 wherein weighting coefficients are generated from the multiple sample matrices and used for setting the characteristics of the ANN.

5. A signal processing system according to claim 1 further including arithmetic means for conducting arithmetic operations on preselected cells of a fixed size matrix generated by the matrix generator to provide an output indicative thereof,

data storage means for storing training data correlated to arithmetic operations conducted by the arithmetic means on multiple sample matrices, and comparator means for comparing the output of the arithmetic means with the training data to verify an entity.

6. A method for signal processing for discriminating time varying band-limited input signals comprising the steps of generating a plurality of time varying band-limited input electrical analog signals such as speech from a plurality of entities to be discriminated;

time encoding each of the input signals into a plurality of time encoded symbols to afford a time encoded signal symbol stream, generating a fixed size matrix composed of a plural coordinate array of cells for receiving each of the time encoded signal symbol streams;

coding each said signal symbol stream into the cells of a fixed size matrix, outputting from the matrices a number of outputs correlated in number with the cells of the fixed size matrix, storing multiple sample matrices indicative of the plurality of entities, and processing the outputs through an artificial neural network processor comprised of an input layer, a hidden layer, and an output layer, including inputting the stored multiple sample matrices into the input layer of the ANN which has a plurality of inputs correlated in number with the outputs from the matrices,

the output layer of the ANN having a plurality of outputs correlated to the entities to be discriminated,

the hidden layer of the ANN having a plurality of inputs and outputs correlated to and receiving the inputs of the input layer and for outputting to the outputs of the output layer, respectively,

the stored multiple sample matrices indicative of each of the plurality of entities setting the characteristics of the ANN to effect discrimination as between the entities.

7. A method of signal processing according to claim 6 further comprising the steps of skeletonizing the outputs of the matrices by employing a parameter of the cell contents of the fixed size matrices to substantially reduce the requirements for inputs to the input layer of the ANN.

8. A method of signal processing according to claim 7 wherein the parameter is a preselected value for cell content whereupon the contents of all cells are discarded which are less than said preselected value.

9. A method of signal processing system according to claim 6 comprising the further steps of generating weighting coefficients from the multiple sample matrices and setting the characteristics of the ANN using value.

10. A method for signal processing system according to claim 6 comprising the further steps of conducting arithmetic operations on preselected cells of the fixed size matrices to provide outputs indicative thereof,

storing training data correlated to the arithmetic operations conducted on multiple sample matrices, and comparing the output of the arithmetic operations with the training data to verify an entity.