Speech recognition in selective call systems

• US 5,719,996 A
• Filed: 06/30/1995
• Issued: 02/17/1998
• Est. Priority Date: 06/30/1995
• Status: Expired due to Fees

• Current Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Cheng, Yan-Ming, Chang, Lu
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/491,329
• Time in Patent Office: 963 Days
• Field of Search: 395/2.65, 395/2.62, 395/2.49
• US Class Current: 704/256
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A selective call communication system (100) has a speech recognition system using an acoustic space (400) which has a plurality of probability density functions (pdfs). The selective call communication system (100) has an acoustic space generator (136) for representing speech in the acoustic space (400) which has a plurality of regions (1-14) having a subset of the plurality of probability density functions (502-516). The selective call communication system (100) has a tree generator (138) for generating a hierarchical tree structure (500) representing the subset of the plurality of probability density functions (502-516) associated with the plurality of regions (1-14), a score computer (132) for determining a region of the plurality of regions (1-14) indicative of a minimum distance to a center of the region for a speech sample received, and a speech recognizer (130) for calculating the probability density functions of the region for recognizing the speech sample received.

Claims

1. A method for performing speech recognition by using an acoustic space having a plurality of probability density functions (pdfs), comprising the steps of:

dividing the acoustic space into a plurality of regions;

determining at least one region of the plurality of regions to which a probability density function belongs by determining when the probability density function has a minimum distance to the at least one of the plurality of regions wherein the probability density functions of each region comprise a subset of the plurality of probability density functions;

generating a hierarchical tree structure representing the subset of the probability density functions associated with the plurality of regions;

receiving a speech sample;

determining the region of the plurality of regions by traversing the hierarchical tree structure in which the speech sample has the minimum distance to a center of the region; and

calculating the subset of probability density functions of the region for recognizing the speech sample received.

2. The method according to claim 1 wherein the step of representing comprises a step of calculating the probability density function using Hidden Markov Model representation.

3. The method according to claim 1 wherein the step of calculating comprises a step of retrieving a list of probability density functions assigned to the region determined to have the minimum distance to the speech sample received.

4. The method of claim 1 wherein the step of determining further comprising a step of using the hierarchical tree structure to determine the minimum distance to the speech sample received.

5. The method of claim 4 wherein the step of using further comprising the steps of:

successively calculating a distance measure at each node of the hierarchical tree structure; and

determining the region of the plurality of regions wherein the speech sample received has a minimum distance from the center of the region.

6. The method according to claim 1 wherein the step of determining further comprises the step of receiving a speech sample corresponding to a pdf.

7. The method according to claim 6 wherein the step of receiving further comprising:

dividing the speech sample received into a multiplicity of speech samples; and

assigning at least one state to each of the multiplicity of speech samples.

8. A method for performing speech recognition by using an acoustic space having a plurality of probability density functions (pdfs), comprising the steps of:

dividing the acoustic space into an at least two regions, the at least two regions;

determining a region of the at least two regions to which a probability density function belongs by determining when the probability density function has a minimum distance from a center of one of the at least two wherein the probability density functions of each region comprises a subset of the plurality of probability density functions;

generating a hierarchical tree structure associated with the at least two regions of the acoustic space wherein the at least two regions having a center comprising a list of pdf of the plurality of pdfs generated by Hidden Markov Model (HMM);

receiving a speech data;

using the hierarchical tree structure for determining a region of the at least two regions in which the received speech data has a minimum distance from the center;

retrieving the list of pdfs associated with the region; and

calculating the pdfs from the list of pdfs for determining a likelihood of the received speech data belonging to the pdfs of the region.

9. The method of claim 8 wherein the step of using further comprising the steps of:

successively calculating a distance measure at each node of the hierarchical tree structure; and

determining the region of the at least two regions wherein the received speech data has a minimum distance from the center of the region.

10. The method according to claim 8 wherein the step of receiving further comprising:

dividing the received speech data into a multiplicity of speech samples; and

assigning at least one state to each of the multiplicity of speech samples.

11. A selective call communication system having speech recognition using an acoustic space comprising a plurality of probability density functions (pdfs), the selective call communication system, comprising:

an acoustic space generator for representing speech in the acoustic space, the acoustic space is divided into a plurality of regions, each of the plurality of regions having a subset of the plurality of probability density functions, the subset of the plurality of probability density functions are associated with a region of the plurality of regions by determining when a probability density function has a minimum distance to the region;

a tree generator for generating a hierarchical tree structure representing the subset of the plurality of probability density functions associated with the plurality of regions;

a score computer for determining the region of the plurality of regions by traversing the hierarchical tree structure in which the speech sample has the minimum distance to a center of the region; and

a speech recognizer for calculating the subset of the plurality of probability density functions of the region for recognizing the speech sample received.

12. The selective call communication system according to claim 11 wherein the speech recognizer calculates the plurality of probability density functions using Hidden Markov Model representation.

13. The selective call communication system according to claim 11 wherein the speech recognizer comprises a hypotheses searcher for retrieving a list of probability density functions assigned to the region determined to have the minimum distance to the speech sample received.

14. The selective call communication system of claim 11 wherein the score computer uses the hierarchical tree structure generated by the tree generator to determine the minimum distance to the speech sample received.

15. The selective call communication system of claim 14 wherein the score computer calculates a distance measure at a node of the hierarchical tree structure for determining the region of the plurality of regions wherein the speech sample received has a minimum distance from the center of the region.

16. The selective call communication system according to claim 11 further comprising a front-end processor for receiving speech samples.

17. The selective call communication system of claim 16 wherein the front-end processor further comprising:

a speech digitizer for dividing the speech samples into a multiplicity of speech samples; and

a feature extractor, coupled to the speech digitizer, for determining characteristic features of the multiplicity of speech samples that are significant to human auditory system.

18. A selective call communication system for performing speech recognition using an acoustic space having a plurality of probability density functions (pdfs), the selective call communication system comprising:

an acoustic space generator for representing speech in the acoustic space and dividing the acoustic space into at least two regions and assigning each region a center, a list of the plurality of probability density functions are associated with a region of the plurality of regions by determining when a probability density function has a minimum distance from the center of the region;

a tree generator for generating a hierarchical tree structure associated with the at least two regions of the acoustic space;

a front-end processor for receiving speech data;

a speech recognizer for using the hierarchical tree structure for determining a region of the at least two regions in which the received speech data has the minimum distance from the center;

a score computer for retrieving the list of the plurality of pdfs associated with the region; and

a hypotheses searcher for calculating pdfs from the list of the plurality of pdfs using the hierarchical tree structure for determining a likelihood of the speech data belonging to the plurality of pdfs of the region.

19. The selective call communication system of claim 18 wherein the speech recognizer determines the region by successively calculating a distance measure at a node of the hierarchical tree structure;

and

the tree generator, responsive to the acoustic space generator, determines the region of the at least two regions wherein the received speech data has a minimum distance from the center of the region.

20. The selective call communication system according to claim 18 wherein the front-end processor further comprising:

a digitizer for dividing the speech data into a multiplicity of speech samples; and

a feature extractor for assigning at least one state to each of the multiplicity of speech samples.