JP3008520B2 - Standard pattern making device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for creating a standard pattern in speech recognition.

[0002]

2. Description of the Related Art Fig. 2 is a collection of the papers of the Acoustical Society of Japan, Spring Research Conference, 1990 2-3-3, "Study on speaker-adaptive unspecified speaker word speech recognition method" (Suzuki, Nakajima; Heisei 2 FIG. 3 is a block diagram showing a conventional standard pattern creation device shown in FIG. In the figure, reference numeral 1 denotes an input terminal to which a voice uttered by a standard speaker is input. Reference numeral 2 denotes an acoustic analysis unit that analyzes a voice input from the input terminal 1 and converts the voice into a feature vector time series. Reference numeral 3 denotes an isometric feature vector time series by performing a normalization process on a time axis with respect to a feature vector time series output from the acoustic analysis means 2 by inputting voices of the same category uttered by a plurality of standard speakers. This is a time axis normalizing means for outputting a group. Reference numeral 4 denotes a dimension extension unit that generates a dimension extension feature vector time series by multiplexing feature vectors at the same point with the isometric feature vector time series group. Reference numeral 5 denotes a dimension extension codebook memory that stores a dimension extension codebook. Reference numeral 6 denotes a dimension extension quantization unit that performs vector quantization on the dimension extension feature vector time series using the dimension extension codebook and converts the time series into a code label time series. Reference numeral 7 denotes a dictionary memory for storing the code label time series. Reference numeral 8 denotes a feature vector extracting unit that extracts a feature vector corresponding to each standard speaker from the dimension-extended codebook and generates a codebook for each standard speaker. Reference numeral 9 denotes a code label by performing vector quantization on the feature vector time series of the standard speaker voice output from the acoustic analysis unit using a codebook for each standard speaker output from the feature vector extraction unit. Vector quantization means for generating a time series and writing the time series in the dictionary memory.

Next, the operation will be described. Here, N
Using standard speakers (where N> 1), category i
(I is an integer from 1 to I, and I indicates the number of categories)
The case where a standard pattern belonging to is created is taken as an example. Also,
It is assumed that the dimension-extended codebook memory has already stored the dimension-extended codebook number 1 (J is a codebook size) created using the voices of the N standard speakers described above.

[0004]

(Equation 1)

[0005] At this time, Vj is a code word with the number j, and is expressed by Equation 2.

[0006]

(Equation 2)

In Equation 2, Equation 3 is

[0008]

(Equation 3)

Speaker number n (n is an integer from 1 to N)
Is an M-dimensional feature vector corresponding to the speech of a standard speaker (hereinafter, referred to as a standard speaker n), and is expressed by Expression 4.

[0010]

(Equation 4)

Therefore, Vj is an M × N-dimensional vector. The method of creating a dimension expansion codebook is described in the Autumn Meeting of the Acoustical Society of Japan, 1988, Fall Meeting, Proceedings 1-32-2, "Applying Speaker Adaptation by Vector Quantization to Continuous Speech Recognition" (Suzuki, Nakajima , Ishikawa; October 1988).

The voice of category i uttered by the standard speaker n is input to the acoustic analysis means 2 through the input terminal 1, and the acoustic analysis means 2 analyzes the input voice to obtain a feature vector time series expressed by the following equation (5). Output.

[0013]

(Equation 5)

Here, Tn, i is the number of streams.

For all standard speakers (n = 1 ... N)
By performing the above-described acoustic analysis processing, a feature vector time series group represented by Expression 6 is created.

[0016]

(Equation 6)

The number of feature vector time series groups 6 is input to the time axis normalizing means 3 and is converted into the number of equal length feature vector time series groups 7 having the same number of series.

[0018]

(Equation 7) here,

[0019]

(Equation 8)

(Where Ti is the number of streams and does not depend on speaker n).

The dimension extending means 4 receives the number 7 of isometric feature vector time series groups as input, and extracts 9 feature vectors at the simultaneous point t for each isometric feature vector time series.

[0022]

(Equation 9)

This is multiplexed to generate the number of dimensionally extended feature vectors 10.

[0024]

(Equation 10)

This is represented by the following equation.

[0026]

[Equation 11]

[Mathematical formula-see original document] where the equation 9 is a feature vector at time t of the isometric vector time series number 8, and is represented by the equation 12.

[0028]

(Equation 12)

Here, Equation 13 is the m-th feature vector number 9
The dimension element, M, is the number of dimensions of the feature vector.

[0030]

(Equation 13)

As a result, the number of dimensions of the dimensionally expanded feature vector becomes M × N. This number of dimensionally expanded feature vectors 1
0 is obtained for t = 1... Ti, and a dimension-extended feature vector time series represented by Expression 14 is output.

[0032]

[Equation 14]

The dimension extension quantization means 6 receives the dimension extension feature vector time series number 14 as an input, and stores a code for minimizing the Euclidean distance from each dimension extension feature vector number 10 in the dimension extension codebook memory 5. A code label corresponding to the code is obtained by searching from the number 1 of the extended dimension codebook, and a code label time series expressed by the expression 15 is generated and written to the dictionary memory 7.

[0034]

(Equation 15)

Here, the expression 16 is a code label.

[0036]

(Equation 16)

As described above, the code label time series number 1
5 is created by vector quantizing the dimensionally extended feature vector obtained by superimposing the feature vectors of the voices of a plurality of standard speakers. It has something to have.

The feature vector extracting means 8 extracts the feature vector of the standard speaker n from the code word Vj of the number of dimension extension codebooks 1 stored in the dimension extension codebook memory 5 in accordance with the relationship of equation 2, The codeword number 4 is obtained. By performing this for j = 1... J, the codebook number 17 of the standard speaker n is generated.

[0039]

[Equation 17]

By performing this process for all standard speakers, a codebook number 18 for all standard speakers is created.

[0041]

(Equation 18)

The vector quantizing means 9 comprises the acoustic analyzing means 2
For the feature vector time series number 5 belonging to the feature vector time series group number 6 which is the output of the above, using the codebook number 17 for the standard speaker n which is the output of the feature vector extraction means 8, A vector quantization process is performed and the number of code label time series is 19

[0043]

[Equation 19]

(However, Expression 20 is a code label, Tn, i
Is the number of series).

[0045]

(Equation 20)

By performing this for n = 1 ... N,
A code label time series group number 21 is generated and written to the dictionary memory.

[0047]

(Equation 21)

As described above, since the generated code label time series is created from the voice of each standard speaker, it has different articulation modes.

With the above operation, the dictionary memory stores N + 1 sets of code label time series having different tonal modalities corresponding to the category i speech.

[0050]

Since the conventional standard pattern creating apparatus is configured as described above, the category i
, One set of code label time series that averages the articulation modes of N standard speakers and N sets of code label time series that are the number of standard speakers according to the articulation mode of each standard speaker. A total of N + 1 sets of code label time series will be generated, and such a standard pattern generation device can be used in a speaker-adaptive speech recognition system in order to improve the recognition rate in response to the variety of articulatory modes of the input speaker. When employed, there is a problem that the number of standard speakers must be increased in order to increase the time series of code labels having different articulation modes.

The present invention has been made in order to solve the above-mentioned problems, and has a standard pattern creating apparatus capable of creating a code label time series having different articulation modes without increasing the number of standard speakers. The purpose is to get.

[0052]

According to the present invention, there is provided a standard pattern creating apparatus comprising: a sound analyzing unit which performs signal analysis by using the same category voice uttered by a plurality of standard speakers as input and converts the signal into a feature vector time series group; Time axis normalization means for normalizing the time axis for the feature vector time series output from the acoustic analysis means, and dimension expansion using the output isometric feature vector time series group output from the time axis normalization means Dimension extension means for generating a feature vector time series, dimension extension codebook memory for storing a dimension extension codebook, and selected for each standard speaker input from the weight data input terminal
Weight data of "0" or more and "1" or less (hereinafter, standard speaker weight)
In accordance with standard speaker weight data of each code of the dimension extension codebook and each dimension extension feature vector of the dimension extension feature vector time series.
Weighted distance calculating means for calculating the calculated distance (hereinafter, referred to as a standard speaker weighted distance), and the standardized speaker weighted distance data obtained by the weighted distance calculating means is used as an input to generate a code label time series. It is provided with a minimum distance code search means for outputting and a dictionary memory for storing the code label time series. Hereinafter, the above-mentioned weighted distance calculation means and minimum distance code search means are collectively referred to as weighted vector quantization means.

[0053]

The weighted vector quantizing means according to the present invention is characterized in that the weighted distance calculating means, which is a constituent element thereof, expands each dimension of the dimensionally extended feature vector time series in accordance with the standard speaker weight data input from the weight data input terminal. The standard speaker weighted distance between the feature vector and each code stored in the dimension expansion codebook memory is calculated, and the minimum distance code searching means, which is also a component of the weighted vector quantization means, With the distance data output from the distance calculation means as input, a code label time series is created by searching for a code that gives the minimum distance to each dimension extended feature vector, and written in the dictionary memory. changing the standard speaker weight data to be input, each time repeating the process in the weighted vector quantization means described above Different code label time series of articulatory manner is generated.

[0054]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. In FIG. 1, 1 is an input terminal, 2 is an acoustic analysis unit, 3 is a time axis normalizing unit, 4 is a dimension expansion unit, 5 is a dimension expansion codebook memory, 7 is a dictionary memory, and the same reference numerals in FIG. The detailed description is omitted because it is the same as or equivalent to the attached conventional device.

Reference numeral 10 denotes a weight data input terminal for inputting standard speaker weight data. Reference numeral 11 denotes each of the feature vectors of the dimension extended feature vector time series output from the dimension extending means 4 and stored in the dimension extended codebook memory 5 according to the standard speaker weight data input from the weight data input terminal 10. Weighted distance calculation means for calculating a weighted distance from each code of the existing dimension expansion codebook. A code label time series 12 is created by using the distance data output from the weighted distance calculation means 11 as an input and searching for a code that gives the minimum distance to each code of the dimension-extended feature vector time series. And means for searching for a minimum distance code to be written into the dictionary memory. Numeral 13 denotes a weighted vector quantization means comprising the weighted distance calculation means 11 and the distance minimum code search means 12.

Next, the operation will be described. Again,
As in the conventional case, the number of standard speakers is N (where N>
It is assumed that 1), and the dimension extension codebook is already stored in the dimension extension codebook memory. Further, a case will be taken as an example where a code label time series is created for a voice of category i and stored in a dictionary memory as in the conventional case. Since the operation from the input of the voice to the dimension expansion means is the same as the conventional operation, the description is omitted, and the description will be made from the weight data input terminal 10.

From the weight data input terminal 10, weight data Wn (where n = 1... N) for each standard speaker n is input. Weight data Wn (where n = 1 ... N)
Determines how much the articulation mode of each standard speaker contributes to the articulation mode of the code label written in the dictionary memory 7 by the weighted vector quantization means 13, and satisfies the equation (22).

[0058]

(Equation 22)

The weighted distance calculating means 11 outputs weight data Wn (however,
n = 1... N), the number of dimension-extended feature vectors 11 of the time-series number of dimension-extended feature vectors 14 corresponding to the category i speech, which is the output of the dimension extending means 4
And the standard speaker weighted distance Di (j, t) between the number of codewords 2 of the number of dimension-extended codebooks 1 (where J is the codebook size) stored in the dimension-extended codebook memory 5 is as follows: Find as in the formula.

[0060]

(Equation 23)

However, Equation 3 is an M-dimensional feature vector as shown in Equation 4, and is obtained by extracting a feature vector corresponding to the standard speaker n from the codeword Vj having the dimension extension codebook number 1. Similarly, Equation 9 is an M-dimensional feature vector as shown in Equation 12, and is obtained by extracting a feature vector corresponding to the standard speaker n from the number of dimensionally expanded feature vectors 10. Equation 24 is the Euclidean distance of the above feature vectors 3 and 9,

[0062]

(Equation 24)

This is obtained by the following equation (25).

[0064]

(Equation 25)

In this way, the standard speaker weighted distance Di
(j, t) is the code word Vj of the dimension extension codebook number 1.
The Euclidean distance between the number of feature vectors 3 corresponding to the standard speaker with the speaker number n extracted from the above and the number of feature vectors 9 of the standard speaker extracted from the number of dimensionally expanded feature vectors 10 is calculated using weight data Wn. the weighted value, because it is obtained by summing the speaker number n = 1 ... N, target
The contribution rate of each standard speaker to the quasi-speaker weighted distance Di (j, t) depends on the weight data Wn (where n = 1... N).
The standard speaker weighted distance Di (j, t) is j = 1 ... J, t
= 1 ... Ti and output as distance data.

The minimum distance code searching means 12 outputs the standard speaker weight as the output of the weighted distance calculating means 11.
Attachment distance data Di (j, t) (where j = 1 ... J, t
= 1 ... Ti), and for any t, the minimum D
By searching for j giving i (j, t) and finding the code label of the code corresponding to j, the code label time series number 2
6

[0067]

(Equation 26)

(However, Expression 27 is a code label) is created and written in the dictionary memory 7.

[0069]

[Equation 27]

Standard input from weight data input terminal 10
By repeating the operations of the above-described weighted distance calculation means 11 and minimum distance code search means 12 while changing the speaker weight data, a plurality of code label time series are created in the dictionary memory.

As described above, the code label time series corresponding to the category i voice is obtained by the weighted distance calculating means 1.
1 is created by the minimum distance code search means 12 with the distance data as the output of 1 as an input, so that it has an articulatory mode in which a plurality of standard speakers are weighted with the weight data Wn and averaged. Since a plurality of code label time series corresponding to the category i voice on the dictionary memory are created according to different weight data,
If such a standard pattern creation device is used in a speaker-adaptive speech recognition system to improve the recognition rate corresponding to the variety of articulatory modes of the input speaker, the standard speech generation method has different articulation modes. Standard patterns with different articulation modes can be created without increasing the number of participants. In the above-described embodiment, the hardware is configured by dedicated hardware. However, the hardware may be realized by software processing in a general-purpose computer or a signal processor.

[0072]

As described above, according to the present invention, sound analysis means for performing signal analysis using the same category voice uttered by a plurality of standard speakers as input and converting the signal into a feature vector time series group, and this sound analysis means Time axis normalizing means for normalizing the time axis for the feature vector time series group which is the output of
Dimension extension means for generating a dimension extension feature vector time series using the isometric feature vector time series output from the time axis normalization means, dimension extension codebook memory for storing dimension extension codebook, and weight data input Weighted distance calculating means for calculating a standard speaker weighted distance between each code of the dimension-extended codebook and each dimension-extended feature vector of the dimension-extended feature vector time series using standard speaker weight data input from a terminal And a minimum distance code search means for outputting a code label time series with the distance data obtained by the weighted distance calculation means as an input, and a dictionary memory for storing the code label time series. A plurality of code label time series with different articulation modes can be created regardless of the number. To identify system, if it is adopted for the deal to diversity recognition rate improvement of articulation aspect of the input speaker, it can be a different standard pattern is created with no articulation manner to increase the standard number of speakers.

[0073]

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a standard pattern creation device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a conventional standard pattern creation device.

[Explanation of symbols]

 Reference Signs List 1 audio signal input terminal 2 acoustic analysis means 3 time axis normalization means 4 dimensional expansion means 5 dimensional expansion codebook memory 6 dimensional expansion quantization means 7 dictionary memory 8 feature vector extraction means 9 vector quantization means 10 weight data input terminal 11 weighted distance calculation means 12 minimum distance code search means 13 weighted vector quantization means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-2-251999 (JP, A) JP-A-61-261799 (JP, A) Suzuki, Nakajima “Speaker-adaptive unspecified speaker word recognition method Suzuki, Nakajima, "Adaptation of Speaker Adaptation Method to Continuous Speech Recognition by Vector Quantization", Sound Lecture, Fall 1988, P43-44 (58) .Cl. ⁷ , DB name) G10L 3/00-9/20 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. An acoustic analysis means for performing an acoustic analysis by using the same category speech uttered by a plurality of standard speakers as an input and outputting a feature vector time series group, and a feature vector time series group output from the acoustic analysis means. On the other hand, a time axis normalizing means for normalizing the time axis, a dimension extending means for generating a dimension extended feature vector using an isometric feature vector time series output from the time axis normalizing means, and a dimension extension code A dimension expansion codebook memory for storing books, and a selection for each standard speaker input from the weight data input terminal
Used are-option "0" or "1" following weight data, with each dimension extension feature vector and each code of the dimension extension codebook the dimensional extension feature vector time series, each target
The distance weighted by the weight data selected for the quasi-speaker
Weighted distance calculating means for calculating the distance; distance minimum code searching means for outputting a code label time series with the distance data obtained by the weighted distance calculating means as input; dictionary memory for storing the code label time series And a standard pattern making device.