Method for recognizing speech using eigenpronunciations

• US 7,113,908 B2
• Filed: 03/05/2002
• Issued: 09/26/2006
• Est. Priority Date: 03/07/2001
• Status: Expired due to Fees

• Current Assignee: Sony Deutschland GmbH (Sony Group Corp.)
• Original Assignee: Sony Deutschland GmbH (Sony Group Corp.)
• Inventors: Kompe, Ralf, Goronzy, Silke
• Primary Examiner(s): Storm, Donald L.
• Application Number:
  US10/090,861
• Publication Number:
  US 20020128820A1
• Time in Patent Office: 1,666 Days
• Field of Search: 704/244, 704/255, 704/277, 704/243, 704/245
• US Class Current: 704/244
• CPC Class Codes: G10L 15/07 to the speaker

Abstract

To increase the recognition rate and quality in a process of recognizing speech an approximative set of pronunciation rules (APR) for a current pronunciation (CP) of a current speaker is determined in a given pronunciation space (PS) and then applied to a current pronunciation lexicon (CL) so as to perform a speaker specific adaptation of said current lexicon (CL).

Claims

1. Method for recognizing speech comprising the steps of:

providing a pronunciation space (PS) including possible pronunciation rules and/or sets thereof, wherein said pronunciation space (PS) includes an approximate set of pronunciation rules (APR) determined and/or generated in accordance with a current pronunciation (CP) of a current speaker, and

providing at least one current lexicon (CL) or a dictionary of pronunciation variants, which is employed for recognition, adapted to said current speaker by applying said approximate set of pronunciation rules (APR) to it, thereby including speaker specific pronunciation variants to said current lexicon (CL);

projecting said current pronunciation (CP) into said pronunciation space (PS);

calculating distance values (d1, . . . ,d4) with respect to Eigenpronunciations (E1, . . . ,E4);

determining a lowest distance value (d4); and

choosing a set of pronunciation rules as said approximate set of pronunciation rules APR which correspond to the Eigenpronunciations (E4) being assigned to said lowest distance value (d4).

2. Method according to claim 1, wherein a step of adapting said current lexicon (CL) is carried out repeatedly, in particular after completed recognition steps and/or obtained recognition results.

3. Method according to claim 1, wherein a step of determining and/or generating said approximate set of pronunciation rules (APR) is carried out repeatedly, so as to iteratively find an approximate set of pronunciation rules (APR) fitting best to said current pronunciation (CP) and/or accent of said current speaker, in particular to consider temporal pronunciation and/or accent variations of said current speaker and/or in particular after completed recognition steps and/or obtained recognition results.

4. Method according to claim 1, wherein said pronunciation space (PS) is generated and/or provided in a pre-processing step, in particular in advance in a recognition process.

5. Method according to claim 1, wherein said pronunciation space (PS) is derived from a plurality and/or limited number of Eigenpronunciations.

6. Method according to claim 5, wherein said Eigenpronunciations are derived from, contain and/or are representative for certain and given pronunciation rules and/or sets thereof, in particular for at least one non-native speaker of at least one target language (TL) with at least one source language (SL) as a mother or native tongue or language of said current speaker.

7. Method according to claim 1, wherein said pronunciation space (PS) is modified during the process of recognition, in particular after completed recognition steps and/or obtained recognition results and/or in particular by modifying one or more Eigenpronunciations.

8. Method according to claim 1, wherein a step of determining and/or generating said approximate set of pronunciation rules (APR) comprises a step of determining a pronunciation-related position of said current speaker in said pronunciation space (PS), in particular in accordance with a current pronunciation (CP) and/or accent of said current speaker.

9. Method according to claim 1, wherein said approximate set of pronunciation rules (APR) is chosen as a given and specific set of pronunciation rules in said pronunciation space (PS), in particular as a given and specific Eigenpronunciation thereof, which is a next neighbor of the speaker'"'"'s current pronunciation (CP), in particular with respect to a pronunciation-related position.

10. Method according to claim 9, wherein a property of being a next neighbor is evaluated by means of a certain given measure or distance function, in particular by an Euclidean distance, in said pronunciation space (PS).

11. Method according to claim 1, wherein said approximate set of pronunciation rules (APR) is chosen as a weighted mixture, superposition and given pronunciation rules, sets, derivatives, and/or components thereof in said pronunciation space (PS), in particular of one or more Eigenpronunciations.

12. Method according to claim 1, wherein said current lexicon (CL) is at least partially based on and/or derived from a starting lexicon (SL) or initial lexicon, in particular on a canonical lexicon essentially containing canonical pronunciation variants of native speakers of a given target language (TL) only and/or in particular in the case of changing to a different and/or new speaker.

13. Method according to claim 1, wherein a step of determining and/or generating said approximate set of pronunciation rules (APR) is at least partially based on and/or derived from a comparison of the current pronunciation (CP) with a canonical pronunciation, in particular with respect to a given utterance, recognition result and/or in particular in the beginning of a recognition session with a different and/or new speaker.

14. Method according to claim 13, wherein said comparison is essentially based on a recognition step using a starting or canonical lexicon (SL) as said current lexicon (CL).

15. Method according to claim 13, wherein for said comparison at least one recognition step is repeated using a phone or phoneme recognizer, so as to yield a sequence of actually uttered phones, phonemes.

16. Method according to claim 13, wherein for said comparison said current pronunciation (CP) of said current speaker is compared to a canonical pronunciation, in particular so as to generate an initial set of pronunciation rules (IR) and/or to locate the pronunciation-related position of said current speaker in said pronunciation space (PS).

17. Method according to claim 1, wherein from current lexicon (CL) recognition related information, pronunciation variants which are not covered by the speaking behavior and/or by the current pronunciation of the current speaker are removed, so as to decrease the amount of data to be evaluated.

18. Method according to claim 1, which is designed for a plurality of source languages (SL) and/or of target languages (TL), in particular with respect to one or more Eigenpronunciations.

19. System for recognizing speech, comprising:

means for providing a pronunciation space (PS) including possible pronunciation rules and/or sets thereof, wherein said pronunciation space (PS) includes an approximate set of pronunciation rules (APR) determined and/or generated in accordance with a current pronunciation (CP) of a current speaker, and

means for providing at least one current lexicon (CL) or a dictionary of pronunciation variants, which is employed for recognition, adapted to said current speaker by applying said approximate set of pronunciation rules (APR) to it, thereby including speaker specific pronunciation variants to said current lexicon (CL);

means for projecting said current pronunciation (CP) into said pronunciation space (PS);

means for calculating distance values (d1, . . . , d4) with respect to Eigenpronunciations (E1, . . . , E4);

means for determining a lowest distance value (d4); and

means for choosing a set of pronunciation rules as said approximate set of pronunciation rules APR which correspond to the Eigenpronunciations (E4) being assigned to said lowest distance value (d4).