Methods and systems for providing speech recognition systems based on speech recordings logs

• US 8,494,853 B1
• Filed: 01/04/2013
• Issued: 07/23/2013
• Est. Priority Date: 01/04/2013
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Weinstein, Eugene, Mengibar, Pedro J. Moreno
• Primary Examiner(s): Chawan, Vijay B
• Application Number:
  US13/734,296
• Time in Patent Office: 200 Days
• Field of Search: 704/235, 704/270.1, 704/270, 704/251, 704/231, 704/277, 704/257, 704/244, 704/255, 704/240, 704/260, 704/275, 704/233, 434/167, 434/308, 715/234
• US Class Current: 704/235
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems
• CPC Class Codes: G10L 2015/0636 Threshold criteria for the updating

Abstract

Examples of methods and systems for providing speech recognition systems based on speech recordings logs are described. In some examples, a method may be performed by a computing device within a system to generate modified data logs to use as a training data set for an acoustic model for a particular language. A device may receive one or more data logs that comprise at least one or more recordings of spoken queries and transcribe the recordings. Based on comparisons, the device may identify any transcriptions that may be indicative of noise and may remove those transcriptions indicative of noise from the data logs. Further, the device may remove unwanted transcriptions from the data logs and the device may provide the modified data logs as a training data set to one or more acoustic models for particular languages.

Claims

1. A method, comprising:

receiving one or more data logs, wherein the one or more data logs comprise at least one or more recordings of spoken queries;

transcribing the one or more recordings of spoken queries;

identifying within transcriptions of the one or more recordings of spoken queries transcriptions having an occurrence exceeding a threshold, wherein the threshold is based on a comparison of the transcriptions with previous transcribed queries;

processing, by a computing device, recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model;

based on a comparison of the processing using the language model with the processing using the acoustic model, identifying, from the one or more data logs, one or more recordings of spoken queries corresponding to transcriptions deemed to be due to noise and a remainder of the one or more recordings of spoken queries;

generating one or more modified data logs including the remainder of the recordings of spoken queries; and

providing the one or more modified data logs and associated transcriptions of the one or more recordings of spoken queries within the one or more modified data logs as a training data set to update one or more acoustic models for particular languages.

2. The method of claim 1, further comprising:

determining a confidence score based on processing recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model; and

removing from the one or more data logs given recordings of spoken queries based on the confidence score.

3. The method of claim 1, further comprising identifying the transcriptions based also on a length of the transcriptions.

4. The method of claim 1, further comprising removing one or more transcriptions of the one or more recordings of spoken queries that include one or more numerical sequences.

5. The method of claim 1, further comprising removing transcriptions comprising a uniform resource locator (URL).

6. The method of claim 1, wherein receiving one or more data logs further comprises receiving up to a threshold amount of data logs from one or more applications.

7. The method of claim 1, wherein receiving one or more data logs further comprises receiving an amount of speech queries based on a gender of a user.

8. A computer readable medium having stored therein instructions, that when executed by a computing device, cause the computing device to perform functions comprising:

receiving one or more data logs, wherein the one or more data logs comprise at least one or more recordings of spoken queries;

transcribing the one or more recordings of spoken queries;

identifying within transcriptions of the one or more recordings of spoken queries transcriptions having an occurrence exceeding a threshold, wherein the threshold is based on a comparison of the transcriptions with previous transcribed queries;

processing recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model;

based on a comparison of the processing using the language model with the processing using the acoustic model, identifying, from the one or more data logs, one or more recordings of spoken queries corresponding to transcriptions deemed to be due to noise and a remainder of the one or more recordings of spoken queries;

generating one or more modified data logs including the remainder of the recordings of spoken queries; and

providing the one or more modified data logs and associated transcriptions of the one or more recordings of spoken queries within the one or more modified data logs as a training data set to update one or more acoustic models for particular languages.

9. The computer readable medium of claim 8, wherein the functions further comprise:

determining a confidence score based on processing recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model; and

removing from the one or more data logs given recordings of spoken queries based on the confidence score.

10. The computer readable medium of claim 8, wherein the functions further comprise identifying the transcriptions based also on a length of the transcriptions.

11. The computer readable medium of claim 8, wherein the functions further comprise removing one or more transcriptions of the one or more recordings of spoken queries containing one or more numerical sequences.

12. The computer readable medium of claim 8, wherein the functions further comprise removing transcriptions comprising a uniform resource locator (URL).

13. The computer readable medium of claim 8, wherein the function of receiving one or more data logs further comprises receiving up to a threshold amount of data logs from one or more applications.

14. The computer readable medium of claim 8, wherein the functions further comprise receiving an amount of speech queries based on a gender of a user.

15. A system, comprising:

at least one processor; and

data storage comprising program instructions executable by the at least one processor to cause the at least one processor to perform functions comprising;

receiving one or more data logs, wherein the one or more data logs comprise at least one or more recordings of spoken queries;

transcribing the one or more recordings of spoken queries;

identifying within transcriptions of the one or more recordings of spoken queries transcriptions having an occurrence exceeding a threshold, wherein the threshold is based on a comparison of the transcriptions with previous transcribed queries;

processing recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model;

based on a comparison of the processing using the language model with the processing using the acoustic model, identifying, from the one or more data logs, one or more recordings of spoken queries corresponding to transcriptions deemed to be due to noise and a remainder of the one or more recordings of spoken queries;

generating one or more modified data logs including the remainder of the recordings of spoken queries; and

providing the one or more modified data logs and associated transcriptions of the one or more recordings of spoken queries within the one or more modified data logs as a training data set to update one or more acoustic models for particular languages.

16. The system of claim 15, wherein the functions further comprise:

determining a confidence score based on processing recordings of spoken queries corresponding to the identified transcriptions using both a language model and an acoustic model; and

removing from the one or more data logs given recordings of spoken queries based on the confidence score.

17. The system of claim 15, wherein the functions further comprise identifying the transcriptions based also on a length of the transcriptions.

18. The system of claim 15, wherein the functions further comprise removing one or more transcriptions of the one or more recordings of spoken queries that include one or more numerical sequences.

19. The system of claim 15, wherein the functions further comprise removing transcriptions comprising a uniform resource locator (URL).

20. The system of claim 15, wherein the functions further comprise receiving up to a threshold amount of data logs from one or more applications.