Mitigating errors in a distributed speech recognition process

• US 8,050,912 B1
• Filed: 11/12/1999
• Issued: 11/01/2011
• Est. Priority Date: 11/13/1998
• Status: Expired due to Term

• Current Assignee: Google Technology Holdings LLC (Alphabet Inc.)
• Original Assignee: Motorola Mobility, Inc. (Lenovo Group Ltd.)
• Inventors: Pearce, David John Benjamin, Gibbs, Jon Alastair
• Primary Examiner(s): Han; Qi
• Application Number:
  US09/830,306
• Time in Patent Office: 4,372 Days
• Field of Search: 704/270.1, 704/231, 704/265, 704/222, 704/239, 704/240, 704/270, 704/251, 704/221, 704/201, 704/233, 714/776, 714/758
• US Class Current: 704/222
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications
• CPC Class Codes: G10L 19/005 Correction of errors induced by the transmission channel, if related to the coding algorithm

Abstract

A method of mitigating errors in a distributed speech recognition process. The method comprises the steps of identifying a group comprising one or more vectors which have undergone a transmission error, and replacing one or more speech recognition parameters in the identified group of vectors. In one embodiment all the speech recognition parameters of each vector of the group are replaced by replacing the whole vectors, and each respective replaced whole vector is replaced by a copy of whichever of the preceding or following vector without error is closest in receipt order to the vector being replaced. In another embodiment determination of which speech recognition parameter or parameters are to be replaced is performed by predicting, from vectors received without error, a predicted value for each speech recognition parameter within said identified group of vectors, and replacing those speech recognition parameters within the identified group of vectors which are outside of a predetermined threshold relative to their respective predicted value. Also described is an apparatus for mitigating errors in a distributed speech recognition process.

Claims

1. A method of mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location;

the method comprising the steps of;

identifying a group comprising one or more of said vectors which have undergone a transmission error; and

replacing one or more speech recognition parameters in the identified group of vectors, wherein said one or more speech recognition parameters in said identified group of vectors are replaced by respective replacement parameters corresponding to copies of one or more corresponding speech recognition parameters from a different vector, corresponding to a different particular sampling time frame, received without error after said identified group of vectors.

2. A method according to claim 1, wherein determination of which speech recognition parameter or parameters are to be replaced is performed by predicting, from vectors received without error, a predicted value for each speech recognition parameter within said identified group of vectors, and replacing those speech recognition parameters within the identified group of vectors which are outside of a predetermined threshold relative to their respective predicted value.

3. A method according to claim 2, wherein if more than a specified number of speech recognition parameters within said identified group of vectors are outside of their respective predetermined thresholds then all the speech recognition parameters of said identified group of vectors are replaced.

4. A method according to claim 1, wherein said step of identifying a group comprising one or more of said vectors which have undergone a transmission error includes a step of predicting respective predicted values for said speech recognition parameters, determining one or more threshold levels relative to the predicted values, and identifying vector groups as having undergone a transmission error responsive to a weighted analysis of how many speech recognition parameters in a vector group are outside of each of said one or more threshold levels.

5. A method according to claim 1, wherein said step of identifying a group comprising one or more of said vectors which have undergone a transmission error includes a step of determining a difference between corresponding speech recognition parameters from different vectors within a vector group, and identifying a vector group having undergone a transmission error responsive to an analysis of how many of said differences are outside of a predetermined threshold level.

6. A method according to claim 1, wherein said speech recognition parameters are transmitted from said first location to said second location over a radio communications link.

7. A method of mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location;

the method comprising the steps of;

identifying a group comprising one or more of said vectors which have undergone a transmission error; and

replacing one or more speech recognition parameters in the identified group of vectors, wherein all the speech recognition parameters of each vector of said group are replaced by replacing the whole vectors, and each respective replaced whole vector is replaced by a copy of whichever of the preceding or following different vector, corresponding to a different particular sampling time frame is received without error and is closest in receipt order to the vector being replaced.

8. A method according to claim 7, wherein a mode of transmission and a mode of error detection are such that said identified group comprises a pair of consecutive vectors, such that the first vector of said pair is replaced by the second vector of a preceding vector without error and the second vector of said pair is replaced by the first vector of a following vector without error.

9. A method of mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location, the method comprising the steps of:

identifying a group comprising one or more of said vectors which have undergone a transmission error; and

replacing one or more speech recognition parameters in the identified group of vectors, wherein said one or more speech recognition parameters in said identified group of vectors are replaced by respective replacement parameters corresponding to one or more speech recognition parameters from a different vector, corresponding to a different particular sampling time frame, received without error after said identified group of vectors;

wherein determination of which speech recognition parameter or parameters are to be replaced is performed by predicting, from vectors received without error, a predicted value for each speech recognition parameter within said identified group of vectors, and replacing those speech recognition parameters within the identified group of vectors which are outside of a predetermined threshold relative to their respective predicted value; and

wherein those speech recognition parameters which are within a predetermined threshold relative to their respective predicted value are compared with a set of reference vectors to find a best match vector from said set of reference vectors, and those speech recognition parameters which are outside of a predetermined threshold relative to their respective predicted value are replaced by corresponding speech recognition parameters from said best match vector.

10. A method according to claim 9, wherein speech recognition parameters from one or more neighbouring vectors are also compared with the set of reference vectors and the best match with respect to a plurality of consecutive reference vectors is chosen.

11. An apparatus for mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location;

the apparatus comprising;

means for identifying a group comprising one or more of said vectors which have undergone a transmission error; and

means for replacing one or more speech recognition parameters in the identified group of vectors, wherein said one or more speech recognition parameters in said identified group of vectors are replaced by respective replacement parameters corresponding to copies of one or more corresponding speech recognition parameters from a different vector, corresponding to a different particular sampling time frame, received without error after said identified group of vectors.

12. An apparatus according to claim 11, wherein determination of which speech recognition parameter or parameters are to be replaced is performed by predicting, from vectors received without error, a predicted value for each speech recognition parameter within said identified group of vectors, and replacing those speech recognition parameters within the identified group of vectors which are outside of a predetermined threshold relative to their respective predicted value.

13. An apparatus according to claim 12, wherein if more than a specified number of speech recognition parameters within said identified group of vectors are outside of their respective predetermined thresholds then all the speech recognition parameters of said identified group of vectors are replaced.

14. An apparatus according to claim 11, wherein said means for identifying a group comprising one or more of said vectors which have undergone a transmission error includes means for predicting respective predicted values for said speech recognition parameters, means for determining one or more threshold levels relative to the predicted values, and means for identifying vector groups as having undergone a transmission error responsive to a weighted analysis of how many speech recognition parameters in a vector group are outside of each of said one or more threshold levels.

15. An apparatus according to claim 11, wherein said means for identifying a group comprising one or more of said vectors which have undergone a transmission error includes means for determining a difference between corresponding speech recognition parameters from different vectors within a vector group, and means for identifying a vector group having undergone a transmission error responsive to an analysis of how many of said differences are outside of a predetermined threshold level.

16. An apparatus according to claim 11, wherein said speech recognition parameters are transmitted from said first location to said second location over a radio communications link.

17. An apparatus for mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location;

the apparatus comprising;

means for identifying a group comprising one or more of said vectors which have undergone a transmission error; and

means for replacing one or more speech recognition parameters in the identified group of vectors, wherein all the speech recognition parameters of each vector of said group are replaced by replacing the whole vectors, and each respective replaced whole vector is replaced by a copy of whichever of the preceding or following different vector, corresponding to a different particular sampling time frame is received without error and is closest in receipt order to the vector being replaced.

18. An apparatus according to claim 17, wherein a mode of transmission and a mode of error detection are such that said identified group comprises a pair of consecutive vectors, such that the first vector of said pair is replaced by the second vector of a preceding vector without error and the second vector of said pair is replaced by the first vector of a following vector without error.

19. An apparatus for mitigating errors in a distributed speech recognition process, the distributed speech recognition process being one in which speech recognition parameters are arranged in vectors, each vector corresponding to a particular sampling time-frame, and said speech recognition parameters are received at a second location having been transmitted from a first location, the apparatus comprising:

means for identifying a group comprising one or more of said vectors which have undergone a transmission error; and

means for replacing one or more speech recognition parameters in the identified group of vectors, wherein said one or more speech recognition parameters in said identified group of vectors are replaced by respective replacement parameters corresponding to one or more speech recognition parameters from a different vector, corresponding to a different particular sampling time frame, received without error after said identified group of vectors;

wherein determination of which speech recognition parameter or parameters are to be replaced is performed by predicting, from vectors received without error, a predicted value for each speech recognition parameter within said identified group of vectors, and replacing those speech recognition parameters within the identified group of vectors which are outside of a predetermined threshold relative to their respective predicted value; and

wherein those speech recognition parameters which are within a predetermined threshold relative to their respective predicted value are compared with a set of reference vectors to find a best match vector from said set of reference vectors, and those speech recognition parameters which are outside of a predetermined threshold relative to their respective predicted value are replaced by corresponding speech recognition parameters from said best match vector.

20. An apparatus according to claim 19, wherein speech recognition parameters from one or more neighbouring vectors are also compared with the set of reference vectors and the best match with respect to a plurality of consecutive reference vectors is chosen.