Method and device for recognising a phonetic sound sequence or character sequence

• US 7,966,177 B2
• Filed: 08/13/2001
• Issued: 06/21/2011
• Est. Priority Date: 08/13/2001
• Status: Expired due to Fees

• Current Assignee: Keep GmbH
• Original Assignee: Keep GmbH
• Inventors: Geiger, Hans
• Primary Examiner(s): Wozniak; James S
• Assistant Examiner(s): He; Jialong
• Application Number:
  US10/486,847
• Publication Number:
  US 20040199389A1
• Time in Patent Office: 3,599 Days
• Field of Search: 704/232, 704/254, 704/257, 704/270, 704/275, 382/229, 382/156, 382/181, 382/186, 382/187, 382/230
• US Class Current: 704/232
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

The invention relates to a method for recognizing a phonetic sound sequence or a character sequence, e.g. according to the ASCII standards, comprising the following steps: a) the sequence is fed to a neural network, b) in said neural network, a sequence of characteristics is formed from the phonetic sequence or character sequence, by taking into consideration phonetic and/or lexical stored information and/or based on a character string sequence (blank characters), c) the character sequence thus formed is compared with a characteristic combination that has a defined statement content, said combination being formed from stored lexical and semantic information, based on the characteristic sequence, d) step c is repeated using new character combinations until, by the reduction of contradictions, a character combination is found that at least largely corresponds with the character sequence, e) the statement content of the character combination with the least number of contradictions is output as the result and/or an action assigned to the statement content is carried out.

Claims

1. A method for recognizing a phonetic sequence of sounds comprising:

a) sending the phonetic sequence of sounds to a neural network implemented on a computer,

b) forming by the neural network a sequence of features from the phonetic sequence taking into account phonetic stored information,

c) selecting a reference combination of features from saved semantical information on the basis of the sequence of features formed in step b), said reference combination of features having a defined meaningful content, and comparing the sequence of features formed in step b) with the reference combination of features,

d) repeating step c) with new reference combinations of features until, by minimizing contradictions, a reference combination of features that substantially agrees with the sequence of features formed in step b) is found, and

e) if the reference combination of features found in step d) matches one of a plurality of predetermined acceptable combinations of features,

then outputting the meaningful content of the reference combination of features found in step d) as a result,

else forming a new sequence of features by the neural network in accordance with step b), and repeating steps c) to e).

2. A method according to claim 1, wherein an earlier iteration of steps c) to e) comprises storing a reference combination of features found in step d) in a temporary memory and a later iteration of steps c) to e) comprises comparing the sequence of features with the stored reference combination of features.

3. A method according to claim 2, comprising assigning the agreeing reference combination of features to a specific statement area or topic area of the temporary memory on the basis of stored lexical information, and using primary features from said statement area or topic areas to form a reference combination of features.

4. A method according to claim 1, comprising, if no agreeing reference combination of features is found in step d), changing a non-agreeing part of the sequence of features on the basis of a combination of features having as many feature agreements as possible, and executing steps c) to e) with the changed sequence of features.

5. A method according to claim 4, comprising changing the non-agreeing part of the sequence of features using a fuzzy logic in a manner that takes account of the similarity and sequence of the meaningful content defined in reference combinations of features, provided such similarity meets a predetermined requirement.

6. A method according to claim 1, wherein step b) comprises generating vectors that are characteristic of the sequence and/or properties of phonetic sequences and step c) comprises comparing the sequence of features formed in step b) with the reference combination of features on the basis of comparison of vectors.

7. A device for detecting a phonetic sequence of sounds, said device comprising:

an input device for inputting speech,

a first memory containing lexical, semantic and phonetic information,

a second memory containing permissible command statements,

a neural network that accepts the phonetic sequence input in the input device, the neural network having a first level that forms a sequence of features from the phonetic sequence, taking into account the information in the first memory, and also having a second level that, using the information from the first memory, forms combinations of features having a defined meaningful content,

a comparison device for comparing the sequence of features formed by the first level of the neural network with the combinations of features formed by the second level of the neural network,

an evaluation device for evaluating contradictions between posited combinations of features with the sequence of features,

a device that, depending on the result of the evaluation, causes the first level of the neural network to form a new sequence of features, and

an output device for displaying the meaningful content of the least contradictory combination of features or for performing a command associated with the meaningful content of the least contradictory combination of features.

8. A device according to claim 7, comprising a fuzzy logic for changing at least part of the sequence of features taking into account the similarity and sequence of features with combinations fo features stored in the first memory, the fuzzy logic having an output side connected to the second level of the neural network.

9. A device according to claim 7, comprising a temporary memory for storing a recognized combination of features, the temporary memory being connected on its output side to the second level of the neural network.

10. A method for recognizing a sequence of characters comprising:

a) sending the sequence of characters to a neural network implemented on a computer,

b) forming by the neural network a sequence of features from the sequence of characters taking into account lexical stored information,

c) selecting reference combination of features from saved lexical information on the basis of the sequence of features formed in step b), said reference combination of features having a defined meaningful content, and comparing the sequence of features formed in step b) with the reference combination of features,

d) repeating step c) with new reference combinations of features until, by minimizing contradictions, a reference combination of features that substantially agrees with the sequence of features formed in step b) is found, and

e) if the reference combination of features found in step (d) matches one of a plurality of predetermined acceptable combinations of features,

then outputting the meaningful content of the reference combination of features found in step (d) as a result,

else forming a new sequence of features by the neural network in accordance with step b), and repeating steps c) to e).

11. A method according to claim 10, wherein an earlier iteration of steps c) to e) comprises storing a reference combination of features found in step d) in a temporary memory and a later iteration of steps c) to e) comprises comparing the sequence of features with the stored reference combination of features.

12. A method according to claim 11, comprising assigning the agreeing reference combination of features to a specific statement area or topic area of the temporary memory on the basis of stored lexical information, and using primary features from said statement area or topic areas to form a reference combination of features.

13. A method according to claim 10, comprising, if no agreeing reference combination of features is found in step d), changing a non-agreeing part of the sequence of features on the basis of a combination of features having as many feature agreements as possible, and executing steps c) to e) with the changed sequence of features.

14. A method according to claim 13, comprising changing the non-agreeing part of the sequence of features using a fuzzy logic in a manner that takes account of the similarity and sequence of the meaningful content defined in reference combinations of features, provided such similarity meets a predetermined requirement.

15. A method according to claim 10, wherein step b) comprises generating vectors that are characteristic of the sequence and step c) comprises comparing the sequence of features formed in step b) with the reference combination of features on the basis of comparison of vectors.

16. A method according to claim 10, wherein in step b), the neural network forms the sequence of features from the sequence of characters taking into account both the lexical stored information and predetermined characters.

17. A device for detecting a sequence of characters, said device comprising:

an input device for inputting characters,

a first memory containing lexical information,

a second memory containing permissible command statements,

a neural network that accepts the sequence of characters input by the input device, the neural network having a first level that forms a sequence of features from the sequence of characters, taking into account the information in the memory and also having a second level that, using the information from the first memory, forms combinations of features having a defined meaningful content,

a comparison device for comparing the sequence of features formed by the first level of the neural network with the combinations of features formed by the second level of the neural network,

an evaluation device for evaluating contradictions of posited combinations of features with the sequence of features,

a device that, depending on the result of the evaluation, causes the first level of the neural network to form a new sequence of features, and

an output device for displaying the meaningful content of the least contradictory combination of features or for performing a command associated with the meaningful content of the least contradictory combination of features.

18. A device according to claim 17, comprising a fuzzy logic for changing at least part of the sequence of features taking into account the similarity and sequence of features with combinations of features stored in the first memory, the fuzzy logic having an output side connected to the second level of the neural network.

19. A device according to claim 17, comprising a temporary memory for storing a recognized combination of features, the temporary memory being connected on its output side to the second level of the neural network.