Speech recognition system employing multiple grammar networks

• US 5,991,720 A
• Filed: 04/16/1997
• Issued: 11/23/1999
• Est. Priority Date: 05/06/1996
• Status: Expired due to Fees

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Junqua, Jean-Claude, Galler, Michael
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): SAX, ROBERT L
• Application Number:
  US08/834,358
• Time in Patent Office: 951 Days
• Field of Search: 704/256, 704/251, 704/275
• US Class Current: 704/256.5
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications
• CPC Class Codes: H04M 2201/40 using speech recognition speech recognition per se G10L15/00
• CPC Class Codes: H04M 3/42204 Arrangements at the exchange for service or number selection by voice at the terminal H04M1/27
• CPC Class Codes: H04M 3/42314 in private branch exchanges
• CPC Class Codes: H04M 3/51 Centralised call answering arrangements requiring operator intervention , e.g. call or contact centers for telemarketing
• CPC Class Codes: H04M 3/527 Centralised call answering arrangements not requiring operator intervention

Abstract

The input speech is segmented using plural grammar networks, including a network that includes a filler model designed to represent noise or extraneous speech. Recognition processing results in plural lists of candidates, each list containing the N-best candidates generated. The lists are then separately aligned with the dictionary of valid names to generate two lists of valid names. The final recognition pass combines these two lists of names into a dynamic grammar and this dynamic grammar may be used to find the best candidate name using Viterbi recognition. A telephone call routing application based on the recognition system selects the best candidate name corresponding to the name spelled by the user, whether the user pronounces the name prior to spelling, or not.

Claims

1. A single pass method of processing acoustic speech data for word recognition, comprising:

processing said acoustic speech data using a recognizer based on a first grammar model to segment said acoustic speech data in a first way and thereby extract a first plurality of recognition candidates;

processing said acoustic speech data using a recognizer based on a second grammar model different that said first grammar model to segment said acoustic speech data in a second way different than said first way and thereby extract a second plurality of recognition candidates;

aligning said first plurality of recognition candidates with a dictionary of predetermined words to generate a first list of word candidates;

aligning said second plurality of recognition candidates with said dictionary of predetermined words to generate a second list of word candidates;

building a dynamic grammar model from said first and second lists of word candidates; and

processing said acoustic speech data using a recognizer based on said dynamic grammar model to extract the recognized word.

2. The method of claim 1 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models.

3. The method of claim 2 wherein said letter models are represented by Hidden Markov Models.

4. The method of claim 2 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models and said second grammar model includes at least one filler model to represent utterances not defined by said letter models.

5. The method of claim 2 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models and said second grammar model includes at least one filler model to represent utterances not defined by said letter models and one silence model to represent a pause in said acoustic speech data.

6. The method of claim 1 wherein said second grammar model defines a letter spotting grammar.

7. The method of claim 1 wherein said first and second grammar models comprise a plurality of different nodes and wherein said first and second plurality of recognition candidates are extracted by a recognition process that scores said nodes according to how closely said acoustic speech data matches said nodes;

and

wherein said first and second plurality of recognition candidates are extracted by selecting those nodes scored as having the closest match to said acoustic data.

8. The method of claim 1 wherein said second grammar model comprises at least one node for representing noise.

9. The method of claim 1 further comprising using said recognized word to route a telephone call.

10. A single pass method of processing acoustic speech data for spelled name recognition, comprising:

processing said acoustic speech data using a recognizer based on a first grammar model to segment said acoustic speech data in a first way and thereby extract a first plurality of letter candidates;

processing said acoustic speech data using a recognizer based on a second grammar model to segment said acoustic speech data in a second way different than said first way and thereby extract a second plurality of letter candidates;

aligning said first plurality of letter candidates with a dictionary of predetermined names to generate a first list of name candidates;

aligning said second plurality of recognition candidates with said dictionary of predetermined words to generate a second list of name candidates; and

processing said acoustic speech data using a recognizer based on said dynamic grammar model to extract the recognized name.

11. The method of claim 10 wherein said first and second grammar models are different.

12. The method of claim 10 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models.

13. The method of claim 12 wherein said letter models are represented by Hidden Markov Models.

14. The method of claim 10 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models and said second grammar model includes at least one filler model to represent utterances not defined by said letter models.

15. The method of claim 10 wherein said first and second grammar models are network models comprising a plurality of interconnected letter models and said second grammar model includes at least one filler model to represent utterances not defined by said letter models and one silence model to represent a pause in said acoustic speech data.

16. The method of claim 10 wherein said second grammar model defines a letter spotting grammar.

17. The method of claim 10 wherein said first and second grammar models comprise a plurality of different nodes and wherein said first and second plurality of recognition candidates are extracted by a recognition process that scores said nodes according to how closely said acoustic speech data matches said nodes;

and

wherein said first and second plurality of recognition candidates are extracted by selecting those nodes scored as having the closest match to said acoustic data.

18. The method of claim 17 wherein said second grammar model comprises at least one node for representing noise.

19. The method of claim 10 further comprising using said recognized word to route a telephone call.

20. A single pass method of processing acoustic speech data for recognition, comprising:

processing said acoustic speech data using a recognizer based on a first grammar network to segment said acoustic speech data in a first way and thereby extract a first plurality of recognition candidates according to a first speech input criteria;

processing said acoustic speech data using a recognizer based on a second grammar network to segment said acoustic speech data on a second way different than said first way and thereby extract a second plurality of recognition candidates according to a second speech input criteria;

transforming said first and second plurality of recognition candidates into transformed candidates based on at least one set of a priori constraints on said acoustic speech data;

making recognition decision based on said transformed candidates.

21. A single pass method of processing acoustic speech data for recognition, comprising:

separately processing said acoustic speech data using different first and second grammar networks that result in different segmentation of said acoustic speech data to extract speech that has utility from speech that does not;

generating a first plurality of recognition candidates using said first grammar network and a second plurality of recognition candidates using said second grammar network;

transforming said first and second plurality of recognition candidates based on at least one set of a priori constraints about the speech that has utility to generate transformed recognition candidates; and

making a recognition decision based on said transformed recognition candidates.