Task-constrained connected speech recognition of propagation of tokens only if valid propagation path is present

• US 5,819,222 A
• Filed: 10/11/1995
• Issued: 10/06/1998
• Est. Priority Date: 03/31/1993
• Status: Expired due to Term

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Smyth, Samuel Gavin, Ringland, Simon Patrick Alexander
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/530,170
• Time in Patent Office: 1,091 Days
• Field of Search: 395/2.68, 395/2.64, 395/2.52, 395/2.63, 395/2.49, 395/2.6, 395/2.5, 395/2.54
• US Class Current: 704/256.1
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A speech recognition system recognizes connected speech using a plurality of vocabulary nodes, at least one of which has an associated signature. In use, partial recognition paths are examined at decision nodes intermediate the beginning and end of the recognition path, each decision node having an associated set of valid accumulated signatures. A token received by a decision node is only propagated if the accumulated signature of that token is one of those in the set of valid accumulated signatures associated with that decision node.

Claims

1. A token-passing speech recognition system for recognising connected speech, the recognition system comprising:

a network for modelling expected input speech, said network having a plurality of vocabulary nodes, at least one of the vocabulary nodes having an associated signature, and

means for examining partial recognition paths at decision nodes intermediate the beginning and end of the recognition path, each decision node having an associated set of valid accumulated signatures, wherein a token received by a decision node is only propagated if the accumulated signature of that token reflecting the path of said token through the network is one of those in the set of valid accumulated signatures associated with that decision node.

2. A speech recognition system as in claim 1 further including:

comparing means for comparing the score and signature of each token, to determine the path through the network with the best match to the input speech signal and those with the next best alternative matches.

3. A speech recognition system as in claim 2 wherein the scores of the tokens are compared at each decision node of the network, only the n top scoring tokens having valid signatures being propagated to the next nodes(s), where n is greater than one.

4. A speech recognition system as in claim 1 wherein the vocabulary nodes are associated with Hidden Markov Models and/or dynamic time warping.

5. A speech recognition system as in claim 1 wherein at least one of the vocabulary nodes of the network is associated with more than one identical word representation model.

6. A speech recognition system as in claim 5 wherein the best scoring path is processed by the first word representation model of a vocabulary node, the next best by the second and so on until either parallel models or incoming paths run out.

7. A speech recognition system as in claim 1 wherein the connected speech deemed to be recognised is determined by tracing the pointers of the tokens back through the network.

8. A speech recognition system as in claim 1 wherein the signatures are compared to determine both the path with the best match to the input speech signal and that with the second best alternative match.

9. A speech recognition system comprising:

a network for modelling expected input speech, the network having a plurality of vocabulary nodes for processing input speech, each vocabulary node being associated with at least one word representation model operable to produce a first parameter indicating the degree of similarity between the input speech and speech deemed to be recognised;

comparing means for comparing the first parameters from distinct word representation models and, in response to this comparison, for determining the parameter to be propagated through the network;

means for determining, at an output of the network, the parameter having the highest degree of similarity and the input speech deemed to be recognised;

means for allocating a signature to at least one of the vocabulary nodes;

means for associating with each first parameter a second parameter representing the accumulated signatures of a path through the network; and

validation means for validating that the second parameter associated with a first parameter is included in a valid set of second parameters, the first parameters being propagated through the network only when the associated second parameter is one of those in a set of valid second parameters.

10. A speech recognition system as in claim 9 wherein the second parameter validation means are operable to validate the second parameter after the input speech is processed by a word representation model of a vocabulary node.

11. A speech recognition system as in claim 9 wherein the speech deemed to be recognised is determined from the second parameter associated with the first parameter available at the output indicating the highest degree of similarity.

12. A speech recognition system as in claim 9 wherein at least one of the vocabulary nodes is associated with more than one identical word representation model.

13. A speech recognition system as in claim 12 wherein the comparing means propagates a maximum of n first parameters having the highest degrees of similarity, n being an integer greater than one.

14. A speech recognition system as in claim 9 adapted to recognise connected strings of spoken numerical digits.

15. A speech recognition system as in claim 9 comprising a programmed digital signal processing chip.

16. A speech recognition system as in claim 9 connected with a telecommunications network.

17. A method of recognising connected input speech, said method comprising.

generating a network for modelling expected input speech, the network having a plurality of vocabulary nodes, each vocabulary node being associated with at least one word representation model capable of producing a first parameter indicating the degree of similarity between the input speech and the speech deemed to be recognised;

comparing the first parameters from distinct word representation models and, in response to this comparison, determining which parameters are to be propagated through the network;

determining, at an output of the network, the parameter having the highest degree of similarity and the input speech deemed to be recognised;

allocating a signature to at least one of the vocabulary nodes; and

associating with the first parameters second parameters representing the accumulated signatures of a path through the network, the first parameters being propagated through the network only when the associated second parameter is one of those in a set of valid second parameters.

18. A method as in claim 17 wherein the second parameters are validated after processing by a word representation model associated with each vocabulary node.

19. A method as in claim 17 wherein a maximum of n first parameters having the highest degrees of similarity are propagated from a common point of the network, n being an integer greater than one.

20. A method as in claim 17 wherein the second parameter associated with a first parameter available at the output having the highest degree of similarity indicates the input speech deemed to be recognised.

21. A method as in claim 17 wherein, when more than one first parameter is available at the output, the first parameters and the associated second parameters available at the output are compared and the first parameters which have the highest degree of similarity and different associated second parameters are retained.

22. A token-passing speech recognition method for recognising connected speech, the recognition method comprising:

modelling expected input speech using a network having a plurality of vocabulary nodes, at least one of the vocabulary nodes having an associated signature, and partial recognition paths at decision nodes intermediate the beginning and end of the recognition path, each decision node having an associated set of valid accumulated signatures,

wherien a token received by a decision node is only propagated if the accumulated signature of that token reflecting the path of said token through the network is one of those in the set of valid accumulated signatures associated with that decision node.

23. A speech recognition method as in claim 22 further including:

comparing the score and signature of each token, to determine the path through the network with the best match to the input speech signal and those with the next best alternative matches.

24. A speech recognition method as in claim 23 wherein the scores of the tokens are compared at each decision node of the network, only the n top scoring tokens having valid signatures being propagated to the next nodes(s), where n is greater than one.

25. A speech recognition method as in claim 22 wherein the vocabulary nodes are associated with Hidden Markov Models and/or dynamic time warping.

26. A speech recognition method as in claim 22 wherein at least one of the vocabulary nodes of the network is associated with more than one identical word representation model.

27. A speech recognition method as in claim 26 wherein the best scoring path is processed by the first word representation model of a vocabulary node, the next best by the second and so on until either parallel models or incoming paths run out.

28. A speech recognition method as in claim 22 wherein the connected speech deemed to be recognised is determined by tracing the pointers of the tokens back through the network.

29. A speech recognition method as in claim 22 wherein the signatures are compared to determine both the path with the best match to the input speech signal and that with the second best alternative match.