Speech recognition pattern adaptation system using tree scheme

• US 6,173,076 B1
• Filed: 02/02/1996
• Issued: 01/09/2001
• Est. Priority Date: 02/03/1995
• Status: Expired due to Term

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Shinoda, Koichi
• Primary Examiner(s): Bella, Matthew C.
• Application Number:
  US08/595,849
• Time in Patent Office: 1,803 Days
• Field of Search: 382/181, 382/186, 382/187, 382/190, 382/195, 382/203, 382/206, 382/224, 382/225, 382/226, 382/228, 382/229, 382/155, 382/156, 382/157, 382/159, 382/160, 382/244 , 382/251, 382/256, 382/259, 704/244, 704/251, 704/256, 704/259
• US Class Current: 382/226
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

Pattern matching means 3 matches categories of a reference pattern stored in reference pattern memory means 2 and input samples of an input pattern produced by input pattern generating means 1. Data statistics estimating means 6 calculates the numbers of input samples in individual nodes of a tree scheme reference pattern stored in the tree scheme reference pattern memory means 4. Node selecting means 6 selects nodes used for adaptation according to the input sample numbers calculated by the data statistics estimating means 5. Adaptation parameter generating means 7 calculates an adaptation parameter in the Nodes selected by the node selecting means 6. Reference pattern generating means 8 produces an adapted reference pattern using the calculated adaptation parameter.

Claims

1. A pattern adaptation system using a distribution tree scheme for adapting a reference pattern comprised a plurality of different categories based on an input pattern as an aggregate of input samples comprising:

(a) input pattern generating means for generating the input pattern;

(b) reference pattern memory means for storing the reference pattern;

(c) pattern matching means for matching the categories of the reference pattern stored in the reference pattern memory means with input samples of the input pattern generated by the input pattern generating means;

(d) distribution tree scheme reference pattern memory means for storing beforehand a distribution tree scheme reference pattern as a reference pattern expressed in a tree scheme;

(e) data amount estimating means for calculating the input sample numbers, in each node of the tree scheme reference pattern stored in the tree scheme reference pattern memory means, by using the correspondence obtained through the pattern matching by the pattern matching means;

(f) node selection means for selecting a node for the adaptation according to the calculated number of the input samples by the data amount estimating means;

(g) adaptation parameter generating means for calculating an adaptation parameter in a node selected by the node selecting means; and

(h) reference pattern generating means for producing an adapted reference pattern by using the adaptation parameter produced by the adaptation parameter generating means and updating the reference pattern.

2. The pattern adaptation system using a tree scheme according to claim 1, wherein the pattern matching means executes probability-based matching, and the data amount estimating means calculates an expected input sample number, the expected input sample number being used in lieu of the input sample number.

3. The pattern adaptation system using a tree scheme according to claim 1, wherein the input pattern used is a time series of feature vectors obtained through analysis of the input speech.

4. The pattern adaptation system using a tree scheme according to claim 3, wherein a hidden Marcov model, in which the output probability distribution is a mixture Gaussian distribution, is used as the tree scheme reference pattern.

5. The pattern adaptation system with a tree scheme according to claim 4, wherein the tree scheme reference pattern is a tree scheme reference pattern produced by taking considerations of output probability distribution mean vectors in individual stages of the hidden Marcov model.

6. The pattern adaptation system with a tree scheme according to claim 1, wherein the tree scheme is based on vocal sound similarities in the vocal sound theory.

7. The pattern adaptation system with a tree scheme according to claim 1, wherein the tree scheme is based on a probability-based tree scheme, in which the degrees to which sub-nodes belong to parent nodes are shown by real numbers of 0 to 1 and the sum of degrees to which a sub-node belongs to a plurality of parent nodes is 1.

8. The pattern adaptation system with a tree scheme according to claim 1, wherein the reference pattern is updated after operation concerning all input speeches has been ended.

9. The pattern adaptation system with a tree scheme according to claim 1, wherein the reference pattern is updated for every input of input speech.

10. The pattern adaptation system with a tree scheme according to claim 1,

wherein the same tree scheme that is used for a speech recognition mode is used for the pattern adaptation system using a tree scheme.

11. A pattern adaptation system, using a distribution tree scheme for adapting a reference pattern comprising a plurality of different categories based on an input pattern as an aggregate of input samples, comprising the steps of:

matching categories of a reference pattern and an input pattern,

calculating a number of input samples in each node of a tree scheme reference pattern,

selecting a node used for adaptation according to the input sample number calculated,

producing an adaptation parameter in the node selected, and

producing an adapted reference pattern and updating the initial reference pattern based on the produced adaptation parameter.

12. A pattern adaptation method utilizing a distribution tree scheme for adapting a reference pattern comprised of a plurality of different categories based upon an input pattern as an aggregate of input samples, comprising the steps of:

(a) generating an input pattern;

(b) storing a reference pattern in a reference pattern memory;

(c) matching the categories of the reference pattern to obtain a reference pattern match, the reference patterns stored in the reference pattern memory with input samples of the input pattern;

(d) storing beforehand a distribution tree scheme reference pattern, in a distribution tree scheme reference pattern memory, as a reference pattern expressed in a tree scheme;

(e) estimating a data amount by calculating the input sample numbers, in each node of the tree scheme reference pattern stored in the tree scheme reference pattern memory, by utilizing the correspondence obtained through the pattern matching;

(f) selecting a node for the adaptation, according to the calculated number of the input samples, by utilizing the results of step (e);

(g) generating an adaptation parameter for calculating an adaptation parameter in the node selected as a result of step (f); and

(h) generating a reference pattern for producing an adapted reference pattern by using the adaptation parameter and updating the reference pattern.

13. A method for obtaining a speech probability comprising the steps of:

(a) calculating a first plurality of distribution probabilities, wherein the plurality of distribution probabilities correspond to a plurality of nodes comprising a distribution tree structure,

(b) calculating a sum of the plurality of distribution probabilities,

(c) calculating a vocal sound unit probability by utilizing the calculated sum,

(d) calculating at least a word probability by utilizing the vocal sound unit probability,

(e) calculating a speech probability by utilizing a set of results of steps (a) through (d),

(f) adapting a reference pattern based upon an input pattern and the results of step (e), and

(g) adapting an input pattern based upon the results of step (e).

14. A method as recited in claim 13 wherein the distribution tree structure comprises a hidden Marcov model, in which the output probability distribution is a mixture Gaussian distribution.

15. A method as recited in claim 14 wherein the distribution tree structure is produced by taking considerations of output probability distribution mean vectors in individual stages of the hidden Marcov model.

16. A method as recited in claim 13 wherein the distribution tree structure is based upon vocal sound similarities in the vocal sound theory.

17. A method as recited in claim 13 wherein the distribution tree structure is based upon a probability-based distribution tree scheme, in which the degrees to which a third plurality of sub-nodes belong to a fourth plurality of parent nodes are shown by real numbers of 0 to 1 and the sum of degrees to which a sub-node belongs to the fourth plurality of parent nodes is 1.

18. A pattern adaptation apparatus, utilizing a distribution tree scheme for adapting a reference pattern comprised of a plurality of different categories based upon an input pattern as an aggregate of input samples, comprising:

(a) an input pattern generating unit for generating the input pattern, wherein

the input pattern utilized comprises a time series of feature vectors obtained by an analysis of an input speech;

(b) a reference pattern memory unit for storing the reference pattern;

(c) a pattern matching unit for matching the categories of the reference pattern stored in the reference pattern memory unit with input samples of the input pattern generated by the input pattern generating unit, wherein

the pattern matching unit executes a probability-based matching;

(d) a distribution tree scheme reference pattern memory unit for storing beforehand a distribution tree scheme reference pattern as a reference pattern expressed in a tree scheme;

(e) a data amount estimating unit for calculating the input sample numbers, in each node of the tree scheme reference pattern stored in the tree scheme reference pattern memory unit, by utilizing the correspondence obtained through the pattern matching by the pattern matching unit;

(f) a node selection unit for selecting a node for the adaptation according to the calculated number of the input samples by the data amount estimating unit, wherein

the tree scheme is based upon a probability-based tree scheme, and wherein the degrees to which sub-nodes belong to parent nodes are represented utilizing real numbers of 0 to 1, and a sum of degrees to which a sub-node belongs to a plurality of parent nodes is 1;

(g) an adaptation parameter generating unit for calculating an adaptation parameter in a node selected by the node selecting unit; and

(h) a reference pattern generating unit for producing an adapted reference pattern by using the adaptation parameter produced by the adaptation parameter generating unit and updating the reference pattern.

19. A pattern adaptation apparatus as recited in claim 18, wherein

the data amount estimating unit calculates an expected input sample number, and wherein the expected input sample number is utilized instead of the input sample number.

20. A pattern adaptation apparatus as recited in claim 18, wherein

a hidden Marcov model, in which an output probability distribution is a mixture Gaussian distribution, is utilized as the reference pattern tree scheme, and wherein

the tree scheme reference pattern is a tree scheme reference pattern produced by taking considerations of output probability distribution mean vectors in individual stages of the hidden Marcov model.