Information processing apparatus for integrating a plurality of feature parameters

• US 6,718,299 B1
• Filed: 01/05/2000
• Issued: 04/06/2004
• Est. Priority Date: 01/07/1999
• Status: Expired due to Fees

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Kondo, Tetsujiro, Yoshiwara, Norifumi
• Primary Examiner(s): To, Doris H.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US09/478,061
• Time in Patent Office: 1,553 Days
• Field of Search: 704/224, 704/243, 704/244, 704/276, 704/201
• US Class Current: 704/224
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit

Abstract

An information processing apparatus includes a feature parameter detector for detecting feature parameters based on a plurality of input data, a normalizer for normalizing the feature parameters detected by the feature parameter detector while maintaining their feature components, and an integration unit for integrating the feature parameters normalized by the normalizer. In the information processing apparatus, feature parameters from a plurality of input data are normalized based on learning normalization coefficients, and distances from each of the normalized feature parameters and to a normal parameter are calculated. Based on the calculated distances, time-series normalization coefficients for performing speech recognition are determined for the feature parameters. Therefore, optimal normalization coefficients for recognizing the feature parameters at each point of time can be obtained.

Claims

1. An information processing apparatus comprising:

a feature parameter detector for detecting feature parameters based on a plurality of different types of input data, each of said different types of input data being obtained independently, at least one detected feature parameter being associated with each of said different types of input data;

a storage unit for storing normalization information;

a normalizer for normalizing the feature parameters detected by said feature parameter detector associated with said different types of input data, using the normalization information stored in said storage unit, and

an integration unit for integrating the feature parameters normalized by the normalizer,

wherein said normalization information in said storage unit is obtained by detecting distances between a normal parameter and each of normalized feature parameters associated with the different types of learning data; and

comparating the detected distances associated with said different types of learning data.

2. An information processing apparatus according to claim 1, further comprising a recognition unit for recognizing that at least one of said plurality of different types of input data corresponds or does not correspond to an object to be recognized.

3. An information processing apparatus according to claim 2, wherein said recognition unit comprises:

a vector quantizer for performing time-series vector-quantization on outputs from said integration unit;

a distance-transition-model storage unit for storing a plurality of distance-transition models; and

a matching unit for performing matching based on distances between the outputs from said integration unit and each distance-transition model.

4. An information processing apparatus according to claim 1, wherein said feature parameter detector detects time-series feature parameters.

5. An information processing apparatus according to claim 4, further comprising a normalization-information storage unit for storing time-series normalization information corresponding to the feature parameters, wherein said normalizer normalizes the feature parameters, based on the normalization information.

6. An information processing apparatus according to claim 5, wherein the normalization information is generated based on the feature parameters by performing learning beforehand.

7. An information processing apparatus according to claim 1, wherein said normalizer comprises:

a first normalizer for performing time-domain normalization on the feature parameters; and

a second normalizer for normalizing the feature parameters normalized by said first normalizer while maintaining the characteristics of the feature parameters.

8. An information processing apparatus according to claim 1, wherein said normalizer normalizes each of the feature parameters, based on time-series normalization information preset based on relationships among the feature parameters.

9. A learning apparatus comprising:

a normalizer for normalizing, based on first normalization information preset for a plurality of different types of time-series input data, each of said different types of input data being obtained independently, feature parameters of the different types of input data, at least one feature parameter being associated with each of said different types of input data;

a detector for detecting a distance between a normal parameter and each of the normalized feature parameters associated with said different types of input data;

a comparator for comparating the detected distances associated with said different types of input data and for outputting the result of the comparation, and

a normalization information generator for generating, based on the result of the comparation, second normalization information for each of the feature parameters.

10. A learning apparatus according to claim 9, wherein the first normalization information is set to be a value for setting the distance between one of the feature parameters and the normal parameter to be equal to the distance between another one of the feature parameters and the normal parameter.

11. A learning apparatus according to claim 10, wherein said normalization information generator generates the second time-series normalization information by providing a larger weight to one parameter having a shorter distance at each point of time.

12. A learning apparatus comprising:

a feature parameter detector for detecting feature parameters based on a plurality of different types of input data, each of said different types of input data being obtained independently, at least one detected feature parameter being associated with each of said different types of input data;

a first normalizer for normalizing the feature parameters detected by the feature parameter detector associated with said different types of input data among the feature parameters;

a second normalizer for normalizing the feature parameters normalized by the first normalizer based on the order thereof;

a matching unit for detecting distances between a normal parameter and each of the normalized feature parameters, normalized by said second normalizer, associated with said different types of input data;

a comparator for comparating the detected distances associated with said different types of input data and for outputting the result of the comparison, and

a normalization-information generator for generating normalization information based on the result of the comparison.

13. A learning apparatus according to claim 12, wherein said first normalizer performs normalization based on normalization information preset for the feature parameters among the feature parameters.

14. An information processing method comprising the steps of:

detecting feature parameters based on a plurality of different types of input data, each of said different types of input data being obtained independently, at least one detected feature parameter being associated with each of said different types of input data;

normalizing the detected feature parameters using normalization information; and

integrating the normalized feature parameters

wherein said normalization information is previously obtained by detecting distances between a normal parameter and each of normalized feature parameters associated with the different types of learning data; and

comparating the detected distances associated with said different types of learning data.

15. An information processing method according to claim 14, further comprising the step of recognizing, based on outputs from the integrating step, that at least one of the input data corresponds or does not correspond to an object be recognized.

16. A learning method comprising the steps of:

normalizing, based on first normalization information preset for a plurality of time-series input data, feature parameters of the different types of input data, each of said different types of input data being obtained independently, at least one feature parameter being associated with each of said different types of input data;

detecting a distance between a normal parameter and each of the normalized feature parameters associated with said different types of input data;

comparating the detected distances associated with said different types of input data and for outputting the result of the comparison; and

generating, based on the result of the comparison, second normalization information for each of the feature parameters.

17. A learning method comprising the steps of:

detecting feature parameters based on a plurality of, different types of input data, each of said different types of input data being obtained independently, at least one detected feature parameter being associated with each of said different types of input data;

normalizing the detected feature parameters among the feature parameters associated with said different types of input data;

further normalizing the normalized feature parameters normalized based on the order thereof;

detecting distances between a normal parameter and each of the normalized feature parameters, normalized by said second normalizer, associated with said different types of input data;

comparating the detected distances associated with said different types of input data, performing a matching process on each of the further normalized feature parameters; and

generating, based on the result of the comparating, normalization information for each of the feature parameters.