Speech recognition apparatus for consumer electronic applications

• US 5,790,754 A
• Filed: 10/21/1994
• Issued: 08/04/1998
• Est. Priority Date: 10/21/1994
• Status: Expired due to Term

• Current Assignee: Sensory, Inc.
• Original Assignee: Sensory, Inc.
• Inventors: Mozer, Forrest S., Mozer, Todd F., Mozer, Michael C.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US08/327,455
• Time in Patent Office: 1,383 Days
• Field of Search: 395/2.09, 395/2.1, 395/2.14, 395/2.15, 395/2.4, 395/2.41, 395/2.57, 395/2.58, 395/2.6, 395/2.62, 395/2.63, 395/21-24
• 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 method for recognizing an utterance spoken by a user comprising the steps of:

capturing the utterance as an input audio signal;

converting the input audio signal to a digitized representation;

using a single pole digital difference filter repeatedly to obtain a plurality of filtered waveforms from the digitized representation, wherein said single pole digital difference filter is in the form;

space="preserve" listing-type="equation">Y(n)=AY(n-1)+BX(n)+CX(n-1);

extracting estimates of a plurality of acoustic parameters from the plurality of filtered waveforms at successive sampling points;

determining an end time of the spoken utterance and a duration of the spoken utterance;

thereafter

time-normalizing said estimates so that the spoken utterance extends over a predetermined number of time intervals; and

analyzing the time-normalized estimates to identify the utterance.

2. The method of claim 1 wherein said step of using a single pole digital difference filter comprises the substeps of:

configuring the single pole digital difference filter using a predetermined set of filter parameters;

applying the digitized representation to an input of the single pole digital difference filter;

thereafter

modifying the set of filter parameters of the single pole digital difference filter to reconfigure the digital difference filter;

directing a previous output signal of the single pole digital difference filter to the input of the single pole digital difference filter to obtain a new output signal of the single pole digital filter; and

repeating said modifying step and said directing step to simulate a bandpass filter having one of said plurality of waveforms as an output as a cascade of successive filter stages.

3. The method of claim 1 wherein said extracting step occurs concurrently with said capturing step.

4. The method of claim 1 wherein said analyzing step is performed using a pretrained neural network.

5. The method of claim 4 wherein said neural network is a feedforward, multilayer, nonlinear classifier.

6. The method of claim 1 wherein said analyzing step comprises the substep of:

assigning the utterance to one of a plurality of categories, said plurality of categories including a plurality of selected word classes and a class that includes all utterances not included in the plurality of selected word classes.

7. The method of claim 4 wherein said neural network operates as a Bayes classifier.

8. The method of claim 4 further comprising the step of:

training the neural network with training data using a numerical optimization procedure.

9. The method of claim 8 wherein said training step comprises the substeps of:

initializing weights of the neural network to be a set of initial values; and

thereafter

varying the neural network weights in response to training data input to the neural network until an optimal weight set is obtained.

10. The method of claim 9 wherein said step of training further comprises the substeps of:

repeating said steps of initializing and varying using different sets of initial values in said initializing step to obtain a plurality of optimal weight sets; and

thereafter

selecting from the plurality of optimal weight sets, an optimal weight set that adjusts the neural network to have a minimum error given the training data as input.

11. The method of claim 1 wherein one of said acoustic parameters is a zero crossing rate of one of said plurality of filtered waveforms.

12. The method of claim 1 wherein one of said acoustic parameters is a power measure of one of said plurality of filtered waveforms.

13. The method of claim 1 wherein one of said acoustic parameters is said duration of the spoken utterance.

14. The method of claim 1 wherein said step of analyzing comprises the substep of:

comparing the utterance to a stored pattern for speaker verification.

18. The method of claim 1 wherein A, B, and C are selected so that no multiplication by one of A, B, or C required for calculating a single term of Y(n) requires more than two shifts and two adds.

15. Apparatus for recognizing an utterance spoken by a user, said apparatus comprising;

an analog-to-digital converter that converts an audio signal into a digitized representation;

a repeatedly accessed single pole digital difference filter that obtains a plurality of filtered waveforms from the digitized representation, wherein said single pole digital difference filter is in the form;

space="preserve" listing-type="equation">Y(n)=AY(n-1)+BX(n)+CX(n-1);

a feature extractor that extracts estimates of a plurality of acoustic parameters from the plurality of filtered waveforms at successive sampling points;

a timer that determines an end time of the spoken utterance and a duration of the spoken utterance;

a time-normalizer that time-normalizes said estimates so that the spoken utterance extends over a predetermined number of time intervals; and

a classifier that analyzes the time-normalized estimates to identify the spoken utterance.

16. The apparatus of claim 15 further comprising a digital difference filter controller that:

configures the digital difference filter using a predetermined set of filter parameters;

applies the digitized representation to an input of the digital difference filter;

modifies the filter parameters of the digital difference filter to reconfigure the digital difference filter;

directs a previous output signal of the digital difference filter to the input of the digital difference filter to obtain a new output signal of the digital filter; and

repeats said modifying step and said directing step to simulate a bandpass filter having one of said plurality of waveforms as an output.

17. The apparatus of claim 15 wherein said classifier is a pretrained neural network.

21. The apparatus of claim 15 wherein A, B, and C are selected so that no multiplication by one of A, B, or C required for calculating a single term of Y(n) requires more than two shifts and two adds.

19. A method for recognizing an utterance spoken by a user comprising the steps of:

capturing the utterance as an input audio signal;

converting the input audio signal to a digitized representation;

using a digital difference filter to obtain a plurality of filtered waveforms from the digitized representation, wherein said single pole digital difference filter is in the form;

space="preserve" listing-type="equation">Y(n)=AY(n-1)+BX(n)+CX(n-1);

extracting, concurrently with said capturing step, estimates of a plurality of acoustic parameters from the plurality of filtered waveforms at successive sampling points, concurrently with said capturing step;

determining an end time of the spoken utterance and a duration of the spoken utterance;

thereafter

time-normalizing said estimates so that the spoken utterance extends over a predetermined number of time intervals; and

analyzing the time-normalized estimates to identify the utterance.

20. Apparatus for recognizing an utterance spoken by a user, said apparatus comprising;

an audio input device that accepts speech input from the user including the utterance and provides an electrical signal responsive to the speech input;

an analog-to-digital converter that converts the electrical signal into a digitized representation;

a single pole digital difference filter that obtains a plurality of filtered waveforms from the digitized representation, wherein said single pole digital difference filter is in the form;

space="preserve" listing-type="equation">Y(n)=AY(n-1)+BX(n)+CX(n-1);

a feature extractor that extracts, estimates of a plurality of acoustic parameters from the plurality of filtered waveforms at successive sampling points;

a timer that determines an end time of the spoken utterance and a duration of the spoken utterance;

a time-normalizer that time-normalizes said estimates so that the spoken utterance extends over a predetermined number of time intervals; and

a classifier that analyzes the time-normalized estimates to identify the spoken utterance.