Model adaptive apparatus for performing adaptation of a model used in pattern recognition considering recentness of a received pattern data

• US 6,920,421 B2
• Filed: 12/26/2000
• Issued: 07/19/2005
• Est. Priority Date: 12/28/1999
• Status: Expired due to Fees

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Pao, Hongchang
• Primary Examiner(s): Ometz, David L.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US09/748,471
• Publication Number:
  US 20010025276A1
• Time in Patent Office: 1,666 Days
• Field of Search: 704/211, 704/252, 704/256
• US Class Current: 704/211
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence

Abstract

In order to improve recognition performance, a no-speech sound model correction section performs an adaptation of a no-speech sound model which is a sound model representing a no-speech state on the basis of input data observed in an interval immediately before a speech recognition interval for the object of speech recognition and the degree of freshness representing the recentness of the input data.

Claims

1. A model adaptive apparatus for performing an adaptation of a model used in pattern recognition which classifies input data in a time series into one of a predetermined number of models, said model adaptive apparatus comprising:

data extraction means for extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and for outputting the data as extracted data; and

model adaptation means for performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data,

wherein said model adaptation means uses, as said freshness, a function in which the value changes in such a manner as to correspond to the time-related position of said extracted data in said predetermined interval.

2. A model adaptive apparatus according to claim 1, wherein said pattern recognition is performed based on a feature distribution in a feature space of said input data.

3. A model adaptive apparatus according to claim 1, further comprising:

feature extraction means fur extracting the features of said input data;

storage means for storing a predetermined number of models into which said input data is to be classified; and

classification means for classifying the features of said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and for outputting the data as extracted data.

4. A model adaptive apparatus according to claim 1, wherein said input data is speech data.

5. A model adaptive apparatus according to claim 4, wherein said predetermined model is a sound model representing noise in an interval which is not a speech segment.

6. A model adaptive apparatus according to claim 1, wherein said function is a monotonically increasing function which increases as time elapses.

7. A model adaptive apparatus according to claim 6, wherein said function is a linear or nonlinear function.

8. A model adaptive apparatus according to claim 6, wherein said function takes discrete values or continuous values.

9. A model adaptive apparatus according to claim 6, wherein said function is a second-order function, a third-order function, or a higher-order function.

10. A model adaptive apparatus according to claim 6, wherein said function is a logarithmic function.

11. A model adaptive apparatus for performing an adaptation of a model used in pattern recognition which classifies input data in a time series into one of a predetermined number of models, said model adaptive apparatus comprising:

data extraction means for extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and for outputting the data as extracted data;

model adaptation means for performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data;

power spectrum analysis means for receiving said extracted data;

noise characteristic calculation means responsive to environmental noise; and

feature distribution parameter calculation means for producing a feature distribution parameter (Z) in response to said power spectrum analysis means and said noise characteristic calculation means.

12. A model adaptive apparatus according claim 11, further comprising:

a plurality of identification function computation means of which one at least receives a no-speech model, said means receiving said feature distribution parameter (Z) and producing in response a respective identification function (G_s(Z), G₁(Z)-G_k(Z)); and

determination mean responsive to said identification functions to produce a recognition result on the basis of a closest match.

13. A model adaptive apparatus for performing an adaptation of a model used in pattern recognition which classifies input data in a time series into one of a predetermined number of models, said model adaptive apparatus comprising:

data extraction means for extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and for outputting the data as extracted data; and

model adaptation means for performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data.

wherein said data extraction means comprise;

framing means having an input for receiving a source of speech and/or environmental noise and for producing in response data frames;

noise observation interval extraction means for extracting a noise vector for a number (M) of flames in a noise observation interval (Tn);

feature extraction means responsive to said noise vector (a) and to an observation vector (a) in a speech recognition interval to produce a feature vector (y); and

no-speech sound model correction means responsive to said noise vector.

14. A pattern recognition apparatus for classifying input data in a time series into one of a predetermined number of models, said pattern recognition apparatus comprising:

feature extraction means for extracting the features of said input data;

storage means for storing said predetermined number of models;

classification means for classifying the features of said input data into one of said predetermined number of models;

data extraction means for extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and for outputting the data as extracted data; and

model adaptation means for performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data,

wherein said model adaptation means uses, as said freshness, a function in which the value changes in such a manner as to correspond to the time-related position of said extracted data in said predetermined interval.

15. A model adaptive method for performing an adaptation of a model used in pattern recognition which classifies input data in a time series into one of a predetermined number of models, said model adaptive method comprising:

a data extraction step of extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and of outputting the data as extracted data; and

a model adaptation step of performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data,

wherein said model adaptation step uses, as said freshness, a function in which the value changes in such a manner as to correspond to the time-related position of said extracted data in said predetermined interval.

16. A recording medium having recorded therein a program for causing a computer to perform an adaptation of a model used in pattern recognition which classifies input data in a time series into one of a predetermined number of models, said program comprising:

a data extraction step of extracting said input data, corresponding to a predetermined model, which is observed in a predetermined interval, and of outputting the data as extracted data; and

a model adaptation step of performing an adaptation of said predetermined model on the basis of the extracted data in said predetermined interval and the degree of freshness representing the recentness of the extracted data,

wherein said model adaptation step uses, as said freshness, a function in which the value changes in such a manner as to correspond to the time-related position of said extracted data in said predetermined interval.