EXTENDED RECOGNITION DICTIONARY LEARNING DEVICE AND SPEECH RECOGNITION SYSTEM

• US 20100023329A1
• Filed: 01/15/2008
• Published: 01/28/2010
• Est. Priority Date: 01/16/2007
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Onishi, Yoshifumi
• Granted Patent: US 8,918,318 B2
• Time in Patent Office: Days
• US Class Current: 704/244
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting

Abstract

Speech recognition of even a speaker who uses a speech recognition system is enabled by using an extended recognition dictionary suited to the speaker without requiring any previous learning using an utterance label corresponding to the speech of the speaker. An extended recognition dictionary learning device includes an utterance variation data calculating section for comparing an acoustic model sequence output from a speech recognition result and an input correct acoustic model sequence to calculate a correspondence between the models as utterance variation data; an utterance variation data classifying section for classifying the calculated utterance variation data into widely appearing utterance variations and unevenly appearing utterance variations; and a recognition dictionary extending section for defining a plurality of utterance variation sets by combining the classified utterance variations and thereby extending the recognition dictionary for each utterance variation set according to the utterance variations included in each utterance variation set. A speech recognition device uses the extended recognition dictionary for each utterance variation set to output a speech recognition result.

Claims

1. An extended recognition dictionary learning device comprising:

utterance variation data calculating section for comparing an acoustic model sequence output from a speech recognition result and an input correct acoustic model sequence to calculate a correspondence between the models as utterance variation data;

utterance variation data classifying section for classifying the calculated utterance variation data into widely appearing utterance variations and unevenly appearing utterance variations; and

recognition dictionary extending section for defining a plurality of utterance variation sets by combining the classified utterance variations and thereby extending the recognition dictionary for each utterance variation set according to the utterance variations included in each utterance variation set.

2. The extended recognition dictionary learning device according to claim 1, wherein

the utterance variation data classifying section includes:

a first calculating section for calculating utterance variations widely appearing in the utterance variation data as an idf utterance variation vector using an idf value of the utterance variation data; and

a second calculating section for clustering the utterance variations unevenly appearing in the utterance variation data using a tfidf value calculated using a tf value of the utterance variation data and the idf value to calculate a cluster utterance variation vector, and

the recognition dictionary extending section constructs a plurality of utterance variation sets by using only utterance variations having a value of the idf utterance variation vector smaller than a predetermined value or by combining the utterance variations having a value of the idf utterance variation vector smaller than a predetermined value and utterance variations having a value of the cluster utterance variation vector larger than a predetermined value.

3. The extended recognition dictionary learning device according to claim 2, wherein

the recognition dictionary extending section constructs the same number of utterance variation sets as the number of clusters by including in each utterance variation set both the utterance variations having a value of the idf utterance variation vector smaller than a predetermined value and utterance variations having a value of the cluster utterance variation vector larger than a predetermined value.

4. The extended recognition dictionary learning device according to claim 3, wherein

the recognition dictionary extending section constructs the number of utterance variation sets larger by one than the number of clusters by further constructing an utterance variation set including the utterance variations having a value of the idf utterance variation vector smaller than a predetermined value in addition to the same number of utterance variation sets as the number of clusters.

5. The extended recognition dictionary learning device according to claim 1, wherein

the recognition dictionary extending section extends the recognition dictionary to construct an extended recognition dictionary for each utterance variation set by adding, to the recognition dictionary, items in which standard utterances included in the recognition dictionary are replaced with utterance variations included in each of the utterance variation sets under a rule that has previously been set as a recognition dictionary extension rule so as to allow utterance variations to be established as speech of a language to be recognized.

6. The extended recognition dictionary learning device according to claim 2, wherein

the first calculating section calculates utterance variations widely appearing in the utterance variation data as the idf utterance variation vector by using the idf value of the utterance variation data represented by idf(X) calculated by the following equations:

$\begin{array}{cc}\mathrm{idf}<br><br>\left(X\right)=\log <br><br>\frac{N}{\mathrm{dnum}<br><br>\left(X\right)}+1& \left[\mathrm{Numeral}<br><br><br><br>1\right]\end{array}$

where the utterance variation is X, the number of speakers is N, and the number of appearances of the utterance variation X in the utterance variation data of each of the N speakers is dnum(X), and

the second calculating section clusters the utterance variations unevenly appearing in the utterance variation data to calculate a cluster utterance variation vector using the tfidf value of the utterance variation data represented by tfidf(X,p) that is calculated using tf(X,p) calculated by the following equation;

$\begin{array}{cc}\mathrm{tf}<br><br>\left(X,p\right)=\frac{\log <br><br>\left(\mathrm{tnum}<br><br>\left(X,p\right)\right)+1}{\log <br><br>\left(\mathrm{frame}<br><br>\left(p\right)\right)}& \left[\mathrm{Numeral}<br><br><br><br>2\right]\end{array}$

where the number of counts of a frame in which an utterance variation X has appeared in the utterance variation data of the speaker p is tnum(X,p) and the number of frames of the learning speech of the speaker p is frame(p) and the idf(X) according to the following equation;

tfidf(X,p)=tf(X,p)idf(X) 

 

[Numeral 3]

7. A speech recognition device comprising

speech recognition section for performing speech recognition for input speech using the recognition dictionary generated for each utterance variation set that has been learned by the extended recognition dictionary learning device as claimed in claim 1.

8. The speech recognition device according to claim 7, wherein

the speech recognition section selects as hypothesis a final recognition result from recognition results obtained for each extended recognition dictionary based on a majority decision method so as to output the final recognition result.

9. A speech recognition system utilizing the extended recognition dictionary learning device as claimed in claim 1.

10. An extended recognition dictionary learning method, comprising:

a step of comparing an acoustic model sequence output from a speech recognition result and an input correct acoustic model sequence to calculate a correspondence between the models;

a step of classifying calculated utterance variation data into widely appearing utterance variations and unevenly appearing utterance variations; and

a step of defining a plurality of utterance variation sets by combining the classified utterance variations and thereby extending the recognition dictionary for each utterance variation set according to the utterance variations included in each utterance variation set.

11. An extended recognition dictionary learning program allowing a computer to execute:

a step of comparing an acoustic model sequence output from a speech recognition result and an input correct acoustic model sequence to calculate a correspondence between the models;

a step of classifying calculated utterance variation data into widely appearing utterance variations and unevenly appearing utterance variations; and

a step of defining a plurality of utterance variation sets by combining the classified utterance variations and thereby extending the recognition dictionary for each utterance variation set according to the utterance variations included in each utterance variation set.