Speech recognition apparatus for consumer electronic applications

• US 6,021,387 A
• Filed: 02/23/1998
• Issued: 02/01/2000
• Est. Priority Date: 10/21/1994
• Status: Expired due to Term

• Current Assignee: Sensory, Inc.
• Original Assignee: Sensory, Inc.
• Inventors: Mozer, Michael C., Mozer, Forrest S., Mozer, Todd F.
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US09/028,205
• Time in Patent Office: 708 Days
• Field of Search: 704/248, 704/231, 704/253, 704/259, 704/213, 704/232
• US Class Current: 704/232
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A spoken word or phrase recognition device. The device does not require a digital signal processor, large RAM, or extensive analog circuitry. The input audio signal is digitized and passed recursively through a digital difference filter to produce a multiplicity of filtered output waveforms. These waveforms are processed in real time by a microprocessor to generate a pattern that is recognized by a neural network pattern classifier that operates in software in the microprocessor. By application of additional techniques, this device has been shown to recognize an unknown speaker saying a digit from zero through nine with an accuracy greater than 99%. Because of the recognition accuracy and cost-effective design, the device may be used in cost sensitive applications such as toys, electronic learning aids, and consumer electronic products.

Claims

1. A neural network for classifying sampled signals into a plurality of classes comprising:

a representation transformer that accepts samples of a selected signal as input, applies a transformation to the input signal to transform the input signal into a representational space, and outputs a transformed signal, wherein the transformation distributes sampled signals belonging to a given class within the representational space in accordance with a predetermined distribution associated with the given class; and

a Bayes classifier that accepts the transformed signal as an input and assigns the input signal to a class with the highest posterior probability.

2. The neural network of claim 1 wherein the predetermined distribution for at least one of the plurality of classes is a multivariate Gaussian distribution.

3. The neural network of claim 1 wherein the predetermined distribution for at least one of said plurality of classes is a multivariate beta distribution.

4. The neural network of claim 1 wherein the predetermined distribution has a covariance matrix that is a diagonal matrix.

5. The neural network of claim 1 wherein weights of the neural network and parameters of a predetermined distribution associated with one of the plurality of classes are obtained by an optimization procedure.

6. The neural network of claim 5 wherein the optimization procedure is a method of steepest descent.

7. The neural network of claim 1 wherein the mean of each predetermined distribution associated with a class is adjustable.

8. The neural network of claim 1 wherein said Bayes classifier accepts prior probabilities for each class as an input.