Method of key-phase detection and verification for flexible speech understanding

• US 5,797,123 A
• Filed: 12/20/1996
• Issued: 08/18/1998
• Est. Priority Date: 10/01/1996
• Status: Expired due to Term

• Current Assignee: Alcatel-Lucent USA, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Kawahara, Tatsuya, Chou, Wu, Juang, Biing-Hwang, Lee, Chin-Hui
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/771,732
• Time in Patent Office: 606 Days
• Field of Search: 704/256, 704/255, 704/254, 704/240, 704/243, 704/231, 704/241, 704/251, 704/253, 704/257, 704/200, 704/236, 704/239, 704/252
• US Class Current: 704/256.5
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

A key-phrase detection and verification method that can be advantageously used to realize understanding of flexible (i.e., unconstrained) speech. A "multiple pass" procedure is applied to a spoken utterance comprising a sequence of words (i.e., a "sentence"). First, a plurality of key-phrases are detected (i.e., recognized) based on a set of phrase sub-grammars which may, for example, be specific to the state of the dialogue. These key-phrases are then verified by assigning confidence measures thereto and comparing these confidence measures to a threshold, resulting in a set of verified key-phrase candidates. Next, the verified key-phrase candidates are connected into sentence hypotheses based upon the confidence measures and predetermined (e.g., task-specific) semantic information. And, finally, one or more of these sentence hypotheses are verified to produce a verified sentence hypothesis and, from that, a resultant understanding of the spoken utterance.

Claims

1. A method for performing speech recognition of a spoken utterance comprising a plurality of words, the method comprising the steps of:

performing key-phrase detection based on one or more phrase sub-grammars to generate a plurality of detected key-phrases, each detected key-phrase comprising a sequence of one or more recognized words;

performing verification on one or more of said detected key-phrases by assigning confidence measures thereto and comparing said confidence measures to one or more threshold values, thereby generating a set of verified key-phrase candidates;

connecting the verified key-phrase candidates to generate one or more sentence hypotheses based upon predetermined semantic information; and

performing verification on one or more of said sentence hypotheses, thereby generating at least one verified sentence hypothesis.

2. The method of claim 1 wherein said one or more phrase sub-grammars are selected from a set of phrase sub-grammars based upon a dialogue state.

3. The method of claim 1 wherein said one or more phrase sub-grammars have been derived based on a training process using a corpus of speech samples.

4. The method of claim 1 wherein the generation of said sentence hypotheses is further based on said confidence measures.

5. The method of claim 1 wherein the detected key-phrases are labeled with semantic tags.

6. The method of claim 5 wherein the generation of said sentence hypotheses is further based on one or more of said semantic tags.

7. The method of claim 6 wherein the generation of said sentence hypotheses is further based on said confidence measures.

8. The method of claim 7 wherein the generation of said sentence hypotheses comprises determining one or more most likely sentence hypotheses based upon said confidence measures, said semantic tags, and said predetermined semantic information.

9. The method of claim 1 wherein the step of performing key-phrase detection is based on a plurality of subword models, and wherein each of said detected key-phrases further comprises a sequence of one or more subwords.

10. The method of claim 9 wherein the subword models comprise Hidden Markov Models.

11. The method of claim 9 wherein the step of performing verification on said detected key-phrases is based on a set of one or more of said subword models and a corresponding set of anti-subword models.

12. The method of claim 11 wherein the subword models and the anti-subword models comprise Hidden Markov Models.

13. The method of claim 1 wherein the step of performing verification on said sentence hypotheses comprises performing an acoustic verification thereupon.

14. The method of claim 1 wherein the step of performing verification on said sentence hypotheses comprises performing a semantic verification thereupon.

15. The method of claim 1 wherein the generating of the at least one verified sentence hypothesis comprises selecting a most likely one of said sentence hypotheses.

16. The method of claim 1 further comprising the step of generating a semantic frame based upon the at least one verified sentence hypothesis.

17. An apparatus for performing speech recognition of a spoken utterance comprising a plurality of words, the apparatus comprising:

a key-phrase detector adapted to generate a plurality of detected key-phrases based on one or more phrase sub-grammars, each detected key-phrase comprising a sequence of one or more recognized words;

a key-phrase verifier applied to one or more of said detected key-phrases, said key-phrase verifier assigning confidence measures to each of said detected key-phrases and comparing said confidence measures to one or more threshold values, thereby generating a set of verified key-phrase candidates;

a sentence hypothesizer adapted to connect the verified key-phrase candidates to generate one or more sentence hypotheses based upon the predetermined semantic information; and

a sentence hypothesis verifier applied to one or more of said sentence hypotheses, thereby generating at least one verified sentence hypothesis.

18. The apparatus of claim 17 wherein said one or more phrase sub-grammars are selected from a set of phrase sub-grammars based upon a dialogue state.

19. The apparatus of claim 17 wherein said one or more phrase sub-grammars have been derived based on a training process using a corpus of speech samples.

20. The apparatus of claim 17 wherein the sentence hypothesizer generates said sentence hypotheses further based on said confidence measures.

21. The apparatus of claim 17 wherein the detected key-phrases are labeled with semantic tags.

22. The apparatus of claim 21 wherein the sentence hypothesizer generates said sentence hypotheses further based on one or more of said semantic tags.

23. The apparatus of claim 22 wherein the sentence hypothesizer generates said sentence hypotheses further based on said confidence measures.

24. The apparatus of claim 23 wherein the sentence hypothesizer determines one or more most likely sentence hypotheses based upon said confidence measures, said semantic tags, and said predetermined semantic information.

25. The apparatus of claim 17 wherein the key-phrase detector operates based on a plurality of subword models, and wherein each of said detected key-phrases further comprises a sequence of one or more subwords.

26. The apparatus of claim 25 wherein the subword models comprise Hidden Markov Models.

27. The apparatus of claim 25 wherein the key-phrase verifier operates based on a set of one or more of said subword models and a corresponding set of anti-subword models.

28. The apparatus of claim 27 wherein the subword models and the anti-subword models comprise Hidden Markov Models.

29. The apparatus of claim 17 wherein the sentence hypothesis verifier performs an acoustic verification on said sentence hypotheses.

30. The apparatus of claim 17 wherein the sentence hypothesis verifier performs a semantic verification on said sentence hypotheses.

31. The apparatus of claim 17 wherein the sentence hypothesis verifier selects a most likely one of said sentence hypotheses.

32. The apparatus of claim 17 further comprising a semantic frame generator operable to generate a semantic frame based upon the at least one verified sentence hypothesis.