Rapid speech recognition adaptation using acoustic input

• US 9,911,411 B2
• Filed: 06/30/2015
• Issued: 03/06/2018
• Est. Priority Date: 06/02/2015
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Connell, II, Jonathan H., Marcheret, Etienne
• Primary Examiner(s): Safaipour, Houshang
• Assistant Examiner(s): Shah, Bharatkumar S
• Application Number:
  US14/755,596
• Publication Number:
  US 20160358601A1
• Time in Patent Office: 980 Days
• Field of Search: 704224
• CPC Class Codes: G10L 15/075 supervised, i.e. under machine guidance
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 17/00 Speaker identification or verification techniques
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

A method includes the following steps. An acoustic input is obtained from a user, including issuing a verbal prompt to the user and receiving the acoustic input from the user in response to the verbal prompt. One or more acoustic representations are obtained, wherein the one or more acoustic representations are generated from a list of expected responses to the issued verbal prompt. The acoustic input from the user is compared to the one or more acoustic representations. One or more speech recognition parameters are adjusted based on the comparison.

Claims

1. A method, comprising:

receiving one or more speech recognition parameters prior to issuing a verbal prompt to a user;

issuing a verbal prompt to the user;

receiving an acoustic input from the user in response to the verbal prompt;

processing one or more sequences of phonemes to obtain one or more acoustic representations, wherein the one or more sequences of phonemes are generated from a list of expected responses to the issued verbal prompt;

comparing the acoustic input from the user to the one or more acoustic representations to determine an acoustic channel characterization and/or speaker class; and

adjusting one or more speech recognition parameters based on the comparison, wherein the adjustment comprises applying feature space mapping to the acoustic input, and further wherein the one or more adjusted speech recognition parameters are used to adjust a speech recognition module of a speech recognition system to use an acoustic model that is consistent with an acoustic channel characterization and/or speaker class so that the selected acoustic model is used for decoding subsequent acoustic input provided by the user as the conversation progresses;

wherein the steps are performed by at least one processor device coupled to a memory.

2. The method of claim 1, wherein the comparing step comprises performing a channel estimation on the acoustic input.

3. The method of claim 1, wherein the comparing step comprises performing a spectral adaptation on the acoustic input.

4. The method of claim 1, wherein the comparing step comprises performing a voice print analysis on the acoustic input.

5. The method of claim 1, wherein the comparing step comprises performing a variable template matching on the acoustic input.

6. The method of claim 1, wherein the adjusting step comprises detecting a class of speaker based on the comparison results.

7. The method of claim 6, wherein the adjusting step further comprises selecting an acoustic model for the detected class of speaker.

8. The method of claim 1, further comprising selecting the issued verbal prompt from a database.

9. The method of claim 1, wherein the processing step comprises generating one or more sequences of voice frequency estimates in a given format.

10. The method of claim 9, wherein the given format is a mel frequency cepstral format.