System and method for hybrid voice recognition

• US 6,836,758 B2
• Filed: 01/09/2001
• Issued: 12/28/2004
• Est. Priority Date: 01/09/2001
• Status: Expired due to Term

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Chang, Chienchung, Jalil, Suhail, Malayath, Narendranath, Huang, William Yee-Ming, Qi, Yingyong, Garudadri, Harinath, DeJaco, Andrew P., Oses, David Puig, Bi, Ning
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): HARPER, V PAUL
• Application Number:
  US09/757,713
• Publication Number:
  US 20020091522A1
• Time in Patent Office: 1,449 Days
• Field of Search: 704/256, 704/255, 704/254, 704/249, 704/248, 704/246, 704/239, 704/231, 600/300, 382/116, 379/88.01
• US Class Current: 704/231
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

A method and system for speech recognition combines different types of engines in order to recognize user-defined digits and control words, predefined digits and control words, and nametags. Speaker-independent engines are combined with speaker-dependent engines. A Hidden Markov Model (HMM) engine is combined with Dynamic Time Warping (DTW) engines.

Claims

1. A voice recognition system, comprising:

an acoustic processor configured to extract speech parameters from a speech segment;

a plurality of different voice recognition engines coupled to the acoustic processor, wherein each voice recognition engine is configured to produce a plurality of hypotheses and a plurality of scores, each score represents a distance from the speech segment to a corresponding hypothesis, and at least one of the voice recognition engines is a nametag engine; and

a decision logic configured to;

receive the plurality of scores from the plurality of different voice recognition engines,

compute a combined score for a subset of the plurality of voice recognition engines, wherein the subset excludes the nametag engine,

determine a best score for the subset and a best score for the nametag engine,

compare the best score for the subset against the best score for the nametag engine.

2. The voice recognition system of claim 1, wherein the decision logic is configured to compute the combined score by:

arranging each of the plurality of scores in an array;

comparing corresponding elements in at least two arrays to create a new array with the lowest scores of the at least two arrays; and

multiplying each array, except the at least two arrays used to create the new array, by a weighting coefficient in order to generate a plurality of weighted arrays, and to combine the plurality of weighted arrays to yield a combined score.

3. The voice recognition system of claim 1, wherein the plurality of different voice recognition engines includes a speaker-independent voice recognition engine.

4. The voice recognition system of claim 3 wherein the plurality of different voice recognition engines includes a speaker-dependent voice recognition engine.

5. The voice recognition system of claim 1, wherein the plurality of different voice recognition engines includes a speaker-dependent voice recognition engine.

6. The voice recognition system of claim 1, wherein the speaker-independent voice recognition engine is a Hidden Markov Model voice recognition engine.

7. The voice recognition system of claim 1, wherein the speaker-independent voice recognition engine is a Dynamic Time Warping voice recognition engine.

8. A method of voice recognition, comprising:

extracting speech parameters from a speech segment;

producing a hypothesis and a corresponding score for each different voice recognition engine of a plurality of different voice recognition engines based on the extracted speech parameters, wherein the score represents a distance from the speech segment to the hypothesis;

computing a minimum score for each of the plurality of different voice recognition engines;

computing a combined score by weighting the minimum scores of the plurality of voice recognition engines; and

using the combined score to select an analysis method to be performed on a hypothesis corresponding to the smallest score.

9. An apparatus to be used for voice recognition, comprising:

means for extracting speech parameters from a speech segment;

means for producing a hypothesis and a corresponding score for each different voice recognition engine of a plurality of different voice recognition engines based on the extracted speech parameters, wherein the score represents a distance from the speech segment to the hypothesis;

means for computing a minimum score for each of the plurality of different voice recognition engines;

means for computing a combined score by weighting the minimum scores of the plurality of voice recognition engine hypothesis; and

means for using the combined score to select an analysis method to be performed on a hypothesis corresponding to the smallest score.

10. A computer readable media embodying a method of voice recognition, the method comprising:

extracting speech parameters from a speech segment;

producing a hypothesis and a corresponding score for each different voice recognition engine of a plurality of different voice recognition engines based on the extracted speech parameters, wherein the score represents a distance from the speech segment to the hypothesis;

computing a minimum score for each of the plurality of different voice recognition engines;

computing a combined score by weighting the minimum scores of the plurality of voice recognition engines; and

using the combined score to select an analysis method to be performed on a hypothesis corresponding to the smallest score.