Synchronized pattern recognition source data processed by manual or automatic means for creation of shared speaker-dependent speech user profile

• US 7,668,718 B2
• Filed: 08/12/2005
• Issued: 02/23/2010
• Est. Priority Date: 07/17/2001
• Status: Expired due to Fees

• Current Assignee: Custom Speech USA, Inc.
• Original Assignee: Custom Speech USA, Inc.
• Inventors: Kahn, Jonathan, Huttinger, Michael C., Demiroglu, Cenk
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US11/203,671
• Publication Number:
  US 20060149558A1
• Time in Patent Office: 1,656 Days
• Field of Search: 704/270
• US Class Current: 704/270
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

An apparatus for transforming data input by dividing the data input into a uniform dataset with one or more data divisions, processing the uniform dataset to produce a first processed dataset with one or more data divisions, processing the uniform dataset to produce a second processed dataset with one or more data divisions, wherein the first and second processed datasets have the same number of data divisions, and editing data selectively within each one of the one or more divisions of the first and second processed dataset. This apparatus has particular utility in error-spotting in processed datasets, and toward training a pattern recognition application, such as speech recognition, to produce more accurate processed datasets.

Claims

1. An apparatus transforming data input a plurality of diverse data sources selected from the group including text, audio, and image, the diverse data sources selected from the group including real-time and recorded, human and mechanically-generated audio, single-speaker and multispeaker, the apparatus comprising:

means for dividing the data input into a uniform dataset with one or more data segments, each uniform dataset being associated with at least one respective data buffer such that each of the respective data buffers would have the same number of data segments given the same data input;

first means for processing of the uniform dataset to produce a first processed uniform dataset with one or more data segments;

second means for processing of the uniform dataset to produce a second processed uniform dataset with one or more data segments, wherein the first and second processed uniform dataset have the same number of data segments; and

means for editing data selectively within each one of the one or more segments of the first and second processed uniform dataset wherein editing means creates a composite output data tile based on data selection from the first and second processed uniform dataset.

2. The apparatus according to claim 1 wherein the edited data is used to transform operations of at least one of the first and second processing means.

3. The apparatus according to claim 2 wherein one of the first and second processing means is a first speech-language computer application, the apparatus further including a speech user profile that contains data selected from the group including at least one of an acoustic model, a language model, and a lexicon.

4. The apparatus according to claim 3 wherein the speech user profile is speaker-dependent.

5. The apparatus according to claim 4 wherein the first speech-language computer application is selected from the group including speech recognition for dictation, speech recognition for command software, interactive voice response for telephony, speaker recognition, audio mining, text to speech, phonetic generation, machine translation, and natural language understanding.

6. The apparatus according to claim 1 wherein one of the first and second processing means is a computer application referencing user profile data selected from the group including an acoustic model, a language model, a lexicon and formatting preferences.

7. The apparatus according to claim 1 further including means for comparing data within each corresponding one of the one or more data segments of the first and second processed uniform dataset to one another.

8. The apparatus according to claim 1 further including distributing means for distributing one or more edited first and second processed uniform datasets as audio tags for utterances, for words, or for word subunits.

9. The apparatus according to claim 1 wherein the dividing means determines divisions based upon a predetermined boundary definition.

10. The apparatus according to claim 9 wherein the dividing means includes a an interface for graphically displaying the amplitude of the data per unit time to assist in the identification of the predetermined boundary definition.

11. The apparatus according to claim 10 wherein the predetermined boundary definition is selected from the group including a predetermined gradient, a discontinuity, a predetermined audio level, and null data.

12. The apparatus according to claim 1 wherein a reference session file is provided, the reference session file including one or more reference segments, the apparatus further comprising:

means comparing the one or more reference segments on a segment by segment basis to each of the one or more data segments in one or both of the first and second processed uniform datasets;

means tor identically numbering each matching reference segment and data segment; and

means for retagging output in session file to sequentially reference data division.

13. The apparatus according to claim 12 wherein the comparing means uses a least differences boundary method, comprising:

creating list of start times for each of the one or more reference segments in the reference session file;

creating list of start times for each of the one or more data segment in each of the one or both of the first and second processed uniform datasets compared;

sequentially establishing least differences between each segment start time from the reference session file and each respective division start tune from the one or both of the first anti second processed uniform datasets compared;

creating new start times and new divisions for each of the one or both of the first and second processed uniform datasets compared; and

retagging data segments of one or both of the respective first and second processed uniform datasets compared with data from reference session file.

14. The apparatus according to claim 1 wherein one of the first and second processing means is a session file editor having a human perceivable interface.

15. The apparatus according to claim 14 wherein the session file editor further includes means for establishing a distribution text.

16. The apparatus according to claim 15 wherein alterations of data within at least one of the first and second uniform datasets may be made using the humanly perceivable interface without affecting the distribution text.

17. The apparatus according to claim 15 wherein the data editor can simultaneously create a verbatim text and the distribution text.

18. The apparatus according to claim 1 wherein the editing means includes means for providing alternate data choices.

19. The apparatus according to claim 1 further includes an audio pronunciation generator that uses a text-to-speech engine for assistance in creating lexical pronunciation for a user profile.

20. The apparatus according to claim 14 wherein each of the session files are selected from either or both human and automatic processes.

21. The apparatus according to claim 1 further including distributing means for distributing one or more edited first and second processed uniform datasets toward training a user profile underlying a computer application selected from the group including speech recognition for dictation, speech recognition for command software, interactive voice response for telephony, speaker recognition, audio mining, text to speech, phonetic generation machine translation, and natural language understanding.

22. The apparatus according to claim 1 wherein the distributing means distributes data toward training a representational model for a computer-aided diagnosis program having pattern recognition capabilities.

23. The apparatus according to claim 1 further includes a speaker-dependent speech processing system.

24. The apparatus according to claim 23 wherein the speaker-dependent speech processing system includes an automatic question generator optimized for the speaker by an iterative selection process.

25. The apparatus according to claim 23 wherein speaker dependent computer application user profile is based upon user specific data.

26. The apparatus according to claim 1 further comprising means for creating an acoustic model from the collected data, the acoustic model creating means having an questions file generator in cases of data sparsity, the creating means comprising:

means for bottom-up clustering;

means for entering parameter values;

means for determining parameter combination processing completion;

means for automatically generating two or more linguistic question sets; and

means for automatically selecting an optimal question set from the automatically generated two or more linguistic question sets.

27. The apparatus according to claim 26 wherein the means for automatically selecting an optimal question set includes using a maximum likelihood based distance measure.

28. The apparatus according to claim 26 further comprising means for running the acoustic model building means with original acoustic model to create one or more accumulator files and means for running the acoustic model building means to combine the acoustic model and accumulator file to create new acoustic model.

29. The apparatus according to claim 28 further including means for tying states.

30. The apparatus according to claim 29 further including means tor untying states.