JPH04298795A - Standard pattern generation device - Google Patents

Abstract

PURPOSE:To obtain the standard pattern generation device which can generate a chord label time series with a different articulation mode without increasing the number of standard speakers as to a standard pattern generation device which generates the chord label time series with the different articulation mode by using plural standard speakers. CONSTITUTION:A weighted distance arithmetic means 11 calculates a weight distance between each of dimensional extended feature vectors of a dimensional extended feature vector time series obtained from a feature vector time series group of voices of plural standard speakers and each of the chords in a dimensional extended chord book according to weight data inputted from a weight data input terminal 10, and then outputs it as distance data. A distance minimum chord search means 12 inputs the distance data outputted by said weighted distance arithmetic means 11 and searches for the chord having the minimum distance to each dimensional extended feature vector to generate and write a chord label time series in a dictionary memory 7.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a standard pattern creation device for speech recognition.

0002

[Prior Art] Figure 2 shows the Proceedings of the 1990 Spring Research Conference of the Acoustical Society of Japan, 2-3-3 "Study of speaker-adaptive speaker-independent word speech recognition system" (Suzuki, Nakajima; 1990 Year 3
FIG. 2 is a block diagram showing a conventional standard pattern creation device shown in FIG. In the figure, reference numeral 1 denotes an input terminal into which voices uttered by a standard speaker are input. Reference numeral 2 denotes an acoustic analysis means that analyzes the audio input from the input terminal 1 and converts it into a time series of feature vectors. 3 performs normalization processing on the time axis on a group of feature vector time series output from the acoustic analysis means 2 by inputting speech of the same category uttered by multiple standard speakers to create an isometric feature vector time series. This is a time axis normalization means that outputs a group. Reference numeral 4 denotes a dimension expansion means for generating a dimension expansion feature vector time series by multiplexing feature vectors at the same point in time with respect to the equal-length feature vector time series group. Reference numeral 5 denotes a dimension expansion codebook memory that stores a dimension expansion codebook. Reference numeral 6 denotes a dimension expansion quantization means that uses the dimension expansion codebook to vector quantize the dimension expansion feature vector time series and converts it into a code label time series. 7 is a dictionary memory that stores the code label time series. Reference numeral 8 denotes a feature vector extracting means for extracting feature vectors corresponding to each standard speaker from the dimensionally expanded codebook and generating a codebook for each standard speaker. 9 is a code label obtained by vector quantizing the feature vector time series of the standard speaker's speech outputted from the acoustic analysis means using the codebook for each standard speaker, which is the output of the feature vector extraction means. A vector quantization means that generates a time series and writes it into the dictionary memory.

Next, the operation will be explained. Here, N
category i using standard speakers (N>1)
(i is an integer from 1 to I, and I indicates the number of categories)
Let us take as an example the case of creating a standard pattern belonging to . Also,
It is assumed that the dimensionally expanded codebook memory has already stored a dimensionally expanded codebook number 1 (J is the codebook size) created using the voices of the aforementioned N standard speakers.

0004

[Math 1]

[0005] At this time, Vj is a code word numbered j, and is expressed by equation 2.

[0006]

[Math 2]

[0007] In equation 2, equation 3 is

[0008]

[Math 3]

Speaker number n (n is an integer from 1 to N)
This is an M-dimensional feature vector corresponding to the voice of a standard speaker (hereinafter referred to as standard speaker n), which is expressed by Equation 4.

0010

[Math 4]

Therefore, Vj is an M×N dimensional vector. For information on how to create a dimensionally expanded codebook, see Proceedings of the 1986 Autumn Research Conference of the Acoustical Society of Japan, 1-3-2.
2 “Application of speaker adaptation method using vector quantization to continuous speech recognition” (Suzuki, Nakajima, Ishikawa; October 1988).

Speech of category i uttered by standard speaker n is input to acoustic analysis means 2 through input terminal 1, and acoustic analysis means 2 acoustically analyzes the input speech and generates a feature vector time series expressed by Equation 5. Output.

[0013]

[Math 5]

[0014] However, Tn,i is the number of sequences.

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

[0016]

[Math 6]

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

[0018]

[Math 7] here,

[0019]

[Math. 8]

(where Ti is the number of sequences and does not depend on speaker n).

The dimension expansion means 4 inputs the number of equal-length feature vector time series groups, 7, and extracts the number of feature vectors, 9, at the same time point t, for each equal-length feature vector time series,

[0022]

[Math. 9]

By multiplexing these, ten dimensionally extended feature vectors are generated.

[0024]

[Math. 10]

[0025] This can be expressed as the following equation.

[0026]

[Math. 11]

However, Equation 9 is a feature vector at time t of equal-length vector time series number 8, and is expressed as Equation 12.

[0028]

[Math. 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]

[Math. 13]

[0031] As a result, the number of dimensions of the dimensionally expanded feature vector becomes M×N dimensions. The number of dimensionally expanded feature vectors is 1
0 to t=1. ．． ．． Ti is calculated, and a dimensionally expanded feature vector time series expressed by Equation 14 is output.

[0032]

[Math. 14]

The dimension expansion quantization means 6 receives the dimension expansion feature vector time series number 14 as input, and stores in the dimension expansion code book memory 5 a code that minimizes the Euclidean distance with each dimension expansion feature vector number 10. A code label time series expressed by Equation 15 is generated by searching among the dimension expansion codebooks number 1 and obtaining a code label corresponding to the code, and writing it into the dictionary memory 7.

[0034]

[Math. 15]

However, number 16 is a code label.

[0036]

[Math. 16]

As mentioned above, the number of code label time series is 1
5 is created by vector quantizing the dimensionally expanded feature vector obtained by superimposing the feature vectors of the voices of multiple standard speakers. It has become something to have.

The feature vector extracting means 8 extracts the feature vector of the standard speaker n from the codeword Vj of the dimensionally extended codebook number 1 stored in the dimensionally extended codebook memory 5 according to the relationship expressed in equation 2, Obtain codeword number 4. This is j=1. ．． ．． By doing this for J, the number of codebooks for standard speaker n is 17.

[0039]

[Math. 17]

By performing this process for all standard speakers, 18 codebooks are created for all standard speakers.

[0041]

[Math. 18]

The vector quantization means 9 is the acoustic analysis means 2
Using the codebook number 17 for standard speaker n, which is the output of the feature vector extraction means 8, for the feature vector time series number 5 belonging to the feature vector time series group number 6, which is the output of Perform vector quantization processing to reduce the number of code label time series to 19

[0043]

[Math. 19]

(However, number 20 is a code label, Tn,
i is the number of sequences).

[0045]

[Math. 20]

[0046] This is set to n=1. ．． ．． By doing this for N, a code label time series group number 21 is generated, and this is written into the dictionary memory.

[0047]

[Math. 21]

As described above, since the generated code label time series groups are created from the voices of each standard speaker, they have different modes of articulation.

Through the above operations, the code label time series having N+1 sets of different articulation modes corresponding to the voices of category i are stored in the dictionary memory.

[0050]

[Problems to be Solved by the Invention] Since the conventional standard pattern creation device is configured as described above,
For the speech of , one set of code label time series is an average of the articulation modes of N standard speakers, and N sets of code label time series according to the articulation modes of each standard speaker are the number of standard speakers. A total of N+1 sets of code label time series will be generated, and such a standard pattern generation device will be used in a speaker-adaptive speech recognition system to improve the recognition rate while dealing with the diversity of articulatory modes of input speakers. If adopted, there was a problem that the number of standard speakers would have to be increased in order to increase the number of code label time series with different articulation modes.

The present invention was made to solve the above-mentioned problems, and provides a standard pattern creation device that can create chord label time series with different modes of articulation without increasing the number of standard speakers. The purpose is to obtain.

[0052]

[Means for Solving the Problems] A standard pattern creation device according to the present invention includes acoustic analysis means for inputting the same category speech uttered by a plurality of standard speakers, performing signal analysis and converting it into a feature vector time series group; For the feature vector time series group output from this acoustic analysis means, a time axis normalization means normalizes the time axis, and the isometric feature vector time series group output from this time axis normalization means is used to calculate the dimension. dimension expansion means for generating an extended feature vector time series; a dimension expansion codebook memory for storing a dimension expansion codebook; and a dimension expansion means for generating each code of the dimension expansion codebook and the dimension expansion according to weight data inputted from a weight data input terminal. A weighted distance calculation means that calculates a weighted distance between the feature vector time series and each dimension extended feature vector, and a distance minimum code that outputs a code label time series using the distance data obtained by the weighted distance calculation means as input. This system is provided with a search means and a dictionary memory for storing this code label time series. Note that hereinafter, the above-mentioned weighted distance calculation means and distance minimum code search means are collectively referred to as weighted vector quantization means.

[0053]

[Operation] In the weighted vector quantization means of the present invention, the weighted distance calculation means, which is a component thereof, calculates each dimension extended feature vector of the dimension extended feature vector time series according to the weight data input from the weight data input terminal. Minimum distance code searching means, which is also a component of the weighted vector quantization means, calculates the weighted distance with each code stored in the dimension expansion codebook memory.
A code label time series is created by inputting the distance data which is the output of the weighted distance calculation means and searching for a code that gives the minimum distance for each dimension extended feature vector, and writing it into the dictionary memory, so that the weight data Each time the weighting data input from the input terminal is changed and the process in the weighted vector quantization means described above is repeated, a chord label time series with a different articulation mode is generated.

[0054]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An 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 means, 3 is a time axis normalization means, 4 is a dimension expansion means, 5 is a dimension expansion codebook memory, and 7 is a dictionary memory.
The detailed explanation will be omitted since these are the same or equivalent parts to the conventional devices with the same reference numerals.

Further, 10 is a weight data input terminal for inputting weight data. Reference numeral 11 denotes each feature vector of the dimension expansion feature vector time series output from the dimension expansion means 4 and the dimension stored in the dimension expansion codebook memory 5 according to the weight data inputted from the weight data input terminal 10. This is a weighted distance calculation means that calculates a weighted distance with each code of the extended codebook. 12 creates a code label time series by using the distance data that is the output of the weighted distance calculation means 11 as input and searching for a code that provides the minimum distance for each code of the dimension-expanded feature vector time series. and minimum distance code searching means for writing into the dictionary memory. 13, the weighted distance calculation means 11 and the distance minimum code search means 1;
This is a weighted vector quantization means consisting of 2 and 2.

Next, the operation will be explained. In this case too,
As in the conventional case, there are N standard speakers (however, N > 1
), and the dimension-extended codebook is already stored in the dimension-extended codebook memory. Further, as in the conventional case, a case will be taken as an example in which a code label time series is created for audio of category i and stored in the dictionary memory. The operations from voice input to the dimension expansion means are the same as conventional ones, so the explanation will be omitted, and the explanation will start from the weight data input terminal 10.

From the weight data input terminal 10, weight data Wn for each standard speaker n (where n=1...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 time series written in the dictionary memory 7 by the weighted vector quantization means 13, and satisfies Equation 22.

[0058]

[Math. 22]

The weighted distance calculation means 11 receives weight data Wn (however,
n=1. ．． ．． N), the number of dimension extended feature vectors corresponding to the voice of category i, which is the output of the dimension expanding means 4, is 11, and the number of dimension extended feature vector time series is 14.
The weighted distance Di(
j, t) is determined as shown below.

[0060]

[Math. 23]

However, Equation 3 is an M-dimensional feature vector as shown in Equation 4, which is a feature vector corresponding to standard speaker n extracted from the codeword Vj of the dimension expansion codebook number 1. Similarly, Equation 9 is an M-dimensional feature vector as shown in Equation 12, which is a feature vector corresponding to standard speaker n extracted from the dimensionally expanded feature vector number 10. Equation 24 is the Euclidean distance between the feature vector number 3 and number 9,

[0062]

[Math. 24]

##EQU25##

[0064]

[Math. 25]

In this way, the weighted distance Di(j, t
) is the number of feature vectors 3 corresponding to the standard speaker assigned the speaker number n extracted from the code word Vj with the number of dimensional expansion codebooks 1, and the number of feature vectors of the standard speaker extracted from the number 10 of dimension expansion feature vectors. The value obtained by weighting the Euclidean distance with the number of feature vectors 9 using the weight data Wn is set to the speaker number n=1. ．． ．． Since it is obtained by summing over N, the contribution rate of each standard speaker to the weighted distance Di (j, t) depends on the weight data Wn (where n=1...N). The weighted distance Di(j, t) is set to j=1. ．． ．． J.
t=1. ．． ．． Ti is determined and output as distance data.

The distance minimum code search means 12 uses the distance data Di(j, t) which is the output of the weighted distance calculation means 11 (where j=1...J, t=1...
．． Ti) as input, for any t, the minimum Di(
By searching for j that gives j, t) and finding the code label of the code corresponding to that j, the number of code label time series is 26.

[0067]

[Math. 26]

(where number 27 is a code label) is created and written into the dictionary memory 7.

[0069]

[Math. 27]

While changing the weight data input from the weight data input terminal 10, the weighted distance calculation means 1 described above
By repeating steps 1 and 12, a plurality of code label time series are created on the dictionary memory.

As described above, the code label time series corresponding to the voice of category i is calculated by the weighted distance calculation means 1.
Since it is created by the minimum distance code search means 12 using the distance data which is the output of 1 as input, it has an articulation mode that is weighted and averaged by weight data Wn for multiple standard speakers. Since multiple code label time series corresponding to voices of category i on the dictionary memory are created according to different weight data,
They have different modes of articulation, and if such a standard pattern creation device is adopted in a speaker-adaptive speech recognition system to improve the recognition rate by dealing with the diversity of articulatory modes of input speakers, it is possible to Standard patterns with different modes of articulation can be created without increasing the number of participants. In addition, although the above-mentioned embodiment shows the configuration using dedicated hardware, it may be realized by software processing in a general-purpose computer or signal processing processor.

[0072]

As described above, according to the present invention, there is provided acoustic analysis means for inputting speech of the same category uttered by a plurality of standard speakers, performing signal analysis and converting it into a time series group of feature vectors, and this acoustic analysis means. a time axis normalization means for normalizing the time axis of the feature vector time series group output from the
dimension expansion means for generating a dimension expansion feature vector time series using the isometric feature vector time series group output from the time axis normalization means; a dimension expansion codebook memory for storing a dimension expansion codebook; and a weight data input. weighted distance calculation means for calculating a weighted distance between each code of the dimension expansion codebook and each dimension expansion feature vector of the dimension expansion feature vector time series according to weight data inputted from a terminal; and the weighted distance calculation means. It consists of a distance minimum code search means that inputs the distance data obtained by the method and outputs a code label time series, and a dictionary memory that stores this code label time series. When code label time series with different articulatory modes can be created and such a standard pattern creation device is adopted in a speaker-adaptive speech recognition system to improve the recognition rate while dealing with the diversity of the articulatory modes of input speakers. , standard patterns with different modes of articulation can be created without increasing the number of standard speakers.

[0073]

[Brief explanation of drawings]

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

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

[Explanation of symbols]

1 Audio signal input terminal 2 Acoustic analysis means 3 Time axis normalization means 4 Dimension expansion means 5 Dimension expansion codebook memory 6 Dimension 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

Claims (1)

[Claims] Claim 1: Acoustic analysis means for inputting the same category speech uttered by a plurality of standard speakers and outputting a feature vector time series group by performing acoustic analysis, and a feature vector time series group output from the acoustic analysis means. On the other hand, a time-axis normalization means for normalizing the time axis, a dimension expansion means for generating a dimension-expanded feature vector time series using a group of isometric feature vector time series output from the time-axis normalization means, A dimension expansion codebook memory that stores an expansion codebook, and weighting of each code of the dimension expansion codebook and each dimension expansion feature vector of the dimension expansion feature vector time series according to weight data input from a weight data input terminal. weighted distance calculation means for calculating a distance; distance minimum code search means for outputting a code label time series using the distance data obtained by the weighted distance calculation means as input; and a dictionary memory for storing the code label time series. A standard pattern creation device equipped with