Systems, methods and articles of manufacture for improving recognition confidence in hypothesized keywords

• US 5,710,864 A
• Filed: 12/29/1994
• Issued: 01/20/1998
• Est. Priority Date: 12/29/1994
• Status: Expired due to Fees

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Lee, Chin-Hui, Rose, Richard Cameron, Juang, Biing-Hwang
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): COLLINS, ALPHONSO
• Application Number:
  US08/366,162
• Time in Patent Office: 1,118 Days
• Field of Search: 395/2.6-2.66, 395/2.5, 395/2.47
• US Class Current: 704/238
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

Systems, methods and articles of manufacture are provided for adjusting the parameters of ones of a plurality of recognition models. The recognition models collectively represent a vocabulary. The recognition models are utilized to identify a known word represented within a received input signal. The received input signal may include within vocabulary and out of vocabulary words. An output signal representing a confidence measure corresponding to the relative accuracy of the identity of the known word is generated. Particular ones of the plurality of recognition models are adjusted as a function of the output signal to improve the confidence measure. The systems, methods and articles of manufacture are preferably implemented in accordance with discriminative techniques, and the adjustment process is used during either a preferred training or a recognition mode.

Claims

1. A speech recognition system comprising:

a memory for storing a plurality of recognition models wherein each one of said recognition models represents a known word within a vocabulary, said plurality of recognition models including keyword recognition models representing keywords and alternate hypothesis recognition models representing words phonetically similar to the keywords; and

a processor having an input for receiving an input signal representing a speech utterance, the processor processing said input signal utilizing ones of said plurality of recognition models to identify a particular known word represented within said input signal, generating an output signal representing a confidence measure corresponding to the relative accuracy of the identity of said particular known word, and adjusting at least one of said alternate hypothesis recognition models as a function of said output signal to improve said confidence measure.

2. The system as set forth in claim 1 wherein said memory further operates to store a plurality of processing unit instructions.

3. The system as set forth in claim 2 wherein said processor includes one or more processing units and is further operable to retrieve and execute selected ones of said processing unit instructions, said selected ones of said processing unit instructions directing said processor to process said input signal utilizing ones of said plurality of recognition models.

4. The system as set forth in claim 1 wherein said processor utilizes discriminative techniques and wherein said processor further produces an indicator signal indicating that said particular known word is one of a particular keyword and a particular phonetically similar word.

5. The system as set forth in claim 4 wherein said processor subtracts a measure of the probability that said particular known word is a keyword from a measure of the probability that said particular known word is a word phonetically similar to a keyword.

6. The system as set forth in claim 1 wherein said memory includes a keyword recognition model database and an alternate hypothesis recognition model database.

7. The system as set forth in claim 1 wherein said processor is operable during at least one of a training mode and a recognition mode.

8. The system as set forth in claim 1 wherein said processor is further operative to produce one or more feature vectors characterizing said received input signal.

9. The system as set forth in claim 1 wherein said input signal is representative of an unknown speech utterance.

10. The system as set forth in claim 9 wherein said processor is further operable to produce an indicator signal having one of a first state indicating said confidence measure associated with said particular known word corresponds to particular threshold, and a second state indicating said confidence measure associated with said particular known word does not correspond to said particular threshold.

11. The system as set forth in claim 1 wherein said plurality of recognition models includes one or more Hidden Markov Models.

12. A method for improving the accuracy of a speech recognition system, said speech recognition system utilizing a plurality of recognition models to identify a particular known word within a received input signal, said received input signal representative of a speech utterance, each one of said recognition models representing different known words within a vocabulary, said plurality of recognition models including keyword recognition models representing keywords and alternate hypothesis recognition models representing words phonetically similar to the keywords, said method comprising the steps of:

identifying said particular known word utilizing ones of said plurality of recognition models;

generating an output signal representing a confidence measure corresponding to the relative accuracy of said identification step; and

adjusting at least one of said alternate hypotheses recognition models as a function of said output signal to improve said confidence measure.

13. The method as set forth in claim 12 wherein said identification step further includes the step of utilizing discriminative techniques.

14. The method as set forth in claim 13 further including the step of indicating whether said particular known word is one of a particular keyword and a particular phonetically similar word.

15. The method as set forth in claim 14 wherein said adjusting step further includes the step of subtracting a measure of the probability that said particular known word is a keyword from a measure of the probability that said particular known word is a word phonetically similar to a keyword.

16. The method as set forth in claim 13 further including the step of storing said keyword recognition models in a keyword recognition model database and said alternate hypothesis recognition models in an alternate hypothesis recognition model database.

17. The method as set forth in claim 12 wherein said method operates during at least one of a training mode and a recognition mode.

18. The method as set forth in claim 12 wherein said identification step is preceded by the step of producing one or more feature vectors characterizing said received input signal.

19. The method as set forth in claim 12 wherein said input signal is representative of an unknown speech utterance, and said method further includes the step of producing an indicator signal having one of a first state indicating said confidence measure associated with said particular known word corresponds to a particular threshold, and a second state indicating said confidence measure associated with said particular known word does not correspond to said particular threshold.

20. The method as set forth in claim 12 wherein said plurality of recognition models includes one or more Markov Hidden Models.

21. A storage medium for storing a plurality of processing system instructions, said plurality of processing system instructions capable of being read and executed by one or more processing units for performing speech recognition, said memory storage device including:

a first instruction set for storing a plurality of recognition models wherein each one of said recognition models represents a different known word within a vocabulary, said plurality of recognition models including keyword recognition models representing keywords and alternate hypothesis recognition models representing words phonetically similar to the keywords;

a second instruction set for storing a received input signal representing a speech utterance;

a third instruction set for utilizing ones of said plurality of recognition models to identify a particular known word represented within said input signal, generating an output signal representing a confidence measure corresponding to the relative accuracy of the identity of said particular known word, and adjusting at least one of said alternate hypothesis recognition models as a function of said output signal to improve said confidence measure.

22. The storage medium as set forth in claim 21 wherein said storage medium is selected from the group consisting of:

a magnetic device operable to utilize patterns of magnetism to store said processing system instructions,

a semiconductor chip operable to utilize on-off static electric charges to store said processing system instructions, and

an optical memory operable to utilize on-off light beams to store said processing system instructions.

23. The storage medium as set forth in claim 21 wherein said third instruction set further utilizes discriminative techniques.

24. The storage medium as set forth in claim 23 wherein said first instruction set stores said keyword recognition models to a keyword recognition model database and said alternate hypothesis recognition models to an alternate hypothesis recognition model database.

25. The storage medium as set forth in claim 23 wherein said third instruction set further indicates whether said particular known word is one of a particular keyword and a particular phonetically similar word.

26. The storage medium as set forth in claim 25 wherein said third instruction set subtracts a measure of the probability that said particular known word is a keyword from a measure of the probability that said particular known word is a word phonetically similar to a keyword.

27. The storage medium as set forth in claim 21 wherein said third instruction set is operable during at least one of a training mode and a recognition mode.

28. The storage medium as set forth in claim 21 wherein said second instruction set produces one or more feature vectors characterizing said received input signal.

29. The storage medium as set forth in claim 21 wherein said input signal is representative of an unknown speech utterance, and said third instruction set produces an indicator signal having one of a first state indicating said confidence measure associated with said particular known word corresponds to a particular threshold, and a second state indicating said confidence measure associated with said particular known word does not correspond to said particular threshold.

30. The storage medium as set forth in claim 21 wherein said plurality of recognition models include one or more Markov Hidden Models.