JP2958922B2 - Voice feature extraction device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech feature value extraction device that extracts a feature value representing a narrow state in a vocal tract.

<Prior Art> Humans can generate various sounds by forming a narrowing in the vocal tract. For example, rubbing noise is generated by creating a narrowing between the tongue and gums or the palate of the tongue and passing a sufficiently elevated exhalation through this narrow gap.
In addition, the plosive sound forms a closure at one of the lips, gums, and palate, and when the intraoral pressure rises to a sufficient degree, the muscle that created the closure loosens and the closure is released, and the clogged air is released at once. Generated by release from the mouth. In addition, nasal noise is generated when a closure is formed in any of the lips, gums, and palate, and the palate sail lowers, opening a passage to the nasal cavity and causing resonance between the nasal cavity and the palate. Vowels are generated by the tongue forming a narrowing in the vocal tract, but since the narrowing is considerably wider than fricatives and plosives, it can be considered virtually open.

Even if roughly divided in this way, the narrow state in the vocal tract includes a closed state, a narrowed state, and an open state, and actually, humans also have a narrowed state, which is a physiological feature of these developments. It is thought that the state is perceived and the voice is identified.

Therefore, it is very important to use a feature amount (hereinafter, referred to as a narrow value) representing a narrow state in the vocal tract as a voice feature amount when recognizing a voice in a voice recognition device or a speech training machine. That is.

Conventionally, the inventors have proposed the following method for generating a narrow value in the vocal tract (Japanese Patent Application No. 63-186352).
That is, the frequency of the input speech is analyzed to obtain a feature vector for each frame, the distance between this feature vector and a standard pattern prepared for each label prepared in advance is calculated, and the label of the closest standard pattern is obtained as described above. And obtains a label sequence corresponding to the input voice. Here, as an example of the above feature vector, a silent state,
There are nasal consonants, buzz bars, vowels, fricatives, qi and plosives.

From the feature vector label adopted as described above, a narrow value in the vocal tract is obtained as follows. That is, if the label is any of a silent state, a nasal consonant, and a buzz bar, it is assumed that a closure has occurred somewhere in the vocal tract in this section. If the label is any one of fricative, plosive, and qi, it is assumed that a constriction is formed somewhere in the vocal tract. Further, if the label is a vowel or a vowel consonant vowel, the vocal tract is assumed to be in an open state without being narrowed. Thus, 3
It decides a narrower value step by step.

<Problem to be Solved by the Invention> However, in the above-described conventional method for generating a narrow value in the vocal tract, when a label sequence for an input voice is obtained, a feature vector obtained by performing frequency analysis on the input voice is prepared in advance. Since the distance from the standard pattern for each label is calculated and the label of the closest standard pattern is adopted as the label of the feature vector obtained above, there are the following problems.

That is, even if the same word is generated, if the speaker changes, the vocal tract length changes, so that the frequency component of the generated speech waveform also changes. Therefore, when matching the standard pattern for each label prepared in advance with the feature vector as described above, the distance to the different standard pattern is closest and the wrong label is adopted, and the narrow value is incorrectly set. There is a problem that it may be determined by

In order to deal with such a problem, a plurality of standard patterns may be prepared for one label so as to support various speakers. However, there is a problem that a large storage capacity is required and the calculation speed is reduced. Further, there is a problem that an increase in the number of candidates at the time of matching leads to erroneous recognition.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech feature amount extraction device capable of generating a narrow value in a vocal tract that does not depend on a change in a frequency component by a speaker.

<Means for Solving the Problems> In order to achieve the above object, the invention according to claim 1 provides a voice feature extraction device that performs frequency analysis of an input voice and extracts a voice feature from an obtained frequency component. Based on the frequency components obtained by the frequency analysis, a discrimination feature extraction unit that determines a discrimination feature value representing the degree of a discrimination feature related to human articulation, and one of the feature amounts from the discrimination feature value According to a rule for generating a narrow value representing the degree of narrowing of a certain vocal tract, a narrowing value generating unit that generates the narrowing value from the discriminating feature value obtained by the discriminating feature extracting unit includes the discriminating feature value. Therefore, a narrower value in the vocal tract that does not depend on the variation of the frequency component by the speaker is generated.

According to a second aspect of the present invention, in the audio feature extracting apparatus according to the first aspect of the present invention, the rule for generating the narrow value is based on at least the discrimination feature values of the previous frame and the current frame. It is a rule that generates a narrow value.

According to a third aspect of the present invention, in the voice feature extraction device for performing frequency analysis of an input voice and extracting a voice feature amount from the obtained frequency component, based on the frequency component obtained by the frequency analysis, A discrimination feature extraction unit that calculates a discrimination feature value that represents the degree of discrimination characteristics related to human articulation, and a rule that generates the narrower value from the discrimination feature value are created by learning using an error backpropagation algorithm. In accordance with the created rule, a narrowing value generating unit that generates the narrowing value from the discriminating feature value obtained by the discriminating feature extracting unit is provided. It is characterized by generating a narrower value in the vocal tract that does not.

According to a fourth aspect of the present invention, there is provided a voice feature extraction apparatus for performing frequency analysis of an input voice by a feature extraction unit and extracting a feature amount of the voice from the obtained frequency component, wherein the feature extraction unit obtains a frequency component. A rule for generating the narrow value from the acoustic parameters extracted at the time is created by learning using an error back propagation algorithm, and the narrow value generation for generating the narrow value from the acoustic parameter according to the created rule. Comprising a unit, from the acoustic parameters,
It is characterized by generating a narrow value in the vocal tract that does not depend on the variation of the frequency component by the speaker.

<Operation> In the invention according to claim 1, when a voice is input, the input voice is subjected to frequency analysis to obtain a frequency component. Then, based on the frequency components obtained by the frequency analysis, a discrimination feature extracting unit obtains a discrimination feature value indicating a degree of a discrimination characteristic related to human articulation. Then, the narrower value is narrowed from the discrimination feature value obtained by the discrimination feature extraction unit according to a rule for generating a narrower value representing the degree of narrowing of the vocal tract, which is one of the speech feature values, from the discrimination feature value. Generated by the value generator.

Therefore, a narrower value in the vocal tract that does not depend on the variation of the frequency component by the speaker is generated from the discrimination feature value.

Further, in the invention according to claim 2, the narrowing value generation unit reliably performs the narrowing of the current frame in accordance with the rule of generating the narrowing value of the current frame based on the discrimination feature values of the previous frame and the current frame. A value is generated.

Further, in the invention according to claim 3, when a voice is input, the input voice is subjected to frequency analysis to obtain a frequency component. Then, based on the frequency components obtained by the frequency analysis, a discrimination feature value indicating a degree of a discrimination characteristic related to human articulation is obtained by a discrimination feature extraction unit. Then, the narrowing value generation unit creates the rule for generating the narrowing value from the discrimination feature value by learning using an error back propagation algorithm, and obtains the rule according to the rule created by the discrimination feature extraction unit. A narrower value is generated from the discrimination feature value obtained.

Therefore, a narrower value in the vocal tract that does not depend on the variation of the frequency component by the speaker is generated from the discrimination feature value.

In the invention according to claim 4, when a voice is input to the feature extraction unit, the input voice is subjected to frequency analysis by the feature extraction unit to obtain a frequency component. Then,
The narrowing value generation unit creates a rule for generating the narrowing value from the acoustic parameters extracted when obtaining the frequency component in the feature extracting unit by learning using an error back propagation algorithm, and creates the rule by itself. , A narrower value is generated from the above acoustic parameters.

Therefore, a narrower value in the vocal tract that does not depend on the variation of the frequency component by the speaker is generated from the acoustic parameters.

<Example> Hereinafter, the present invention will be described in detail with reference to an illustrated example.

As a feature amount that can well express the difference in the narrow position of the narrow state in the vocal tract without being affected by the frequency fluctuation by the speaker, there is a discrimination feature that indicates a discriminative feature related to human articulation. The present invention is to generate a narrow value in the vocal tract by a value representing the degree of the discrimination feature (hereinafter, referred to as a discrimination feature value).

FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention. The audio signal input from the microphone 1 is amplified by the amplifier 2 and input to a feature extraction unit 3 composed of, for example, a 16-channel BPF (band filter) group. Then, the feature amount such as the frequency component obtained by performing the frequency analysis by the feature extracting unit 3 is input to the discrimination feature extracting unit 4 according to the present invention. The discrimination feature extraction unit 4 is a unit that extracts the discrimination feature value for each frame from the input frequency components by a predetermined procedure.

The time series of the discrimination feature values of the input speech obtained for each frame by the discrimination feature extraction unit 4 is input to the narrowing value generation unit 5. Then, the narrowing value generating unit 5 generates a narrowing value for each frame of the input voice as described in detail later, and the time series of the feature amount of the input voice including the generated narrowing value is a word. It is input to the recognition unit 5. And
The similarity between the feature amount of the standard pattern stored in the standard pattern storage unit 6 and the time series is calculated. At this time, the similarity calculation is performed by a DP matching method or the like. Then, the speech input word recognized based on the result of the similarity calculation is displayed on the result display unit 7.

The narrowing value generator 5 generates a narrowing value by any one of the following three methods. That is, the first method is a method of generating a narrower value from the discrimination feature value according to a rule created based on knowledge,
The second method is a method of automatically creating a rule for generating a narrow value using a neural network described in detail later and generating a narrow value from the discrimination feature value. This is a method of automatically creating a rule for generating a narrow value using the above-described neural network, and directly generating a narrow value from acoustic parameters extracted when a frequency component is obtained by the feature extracting unit 3. .

Here, the discrimination features used in the present embodiment include “strident” indicating friction, “nasal” indicating nasal voice, “murmur / buzz” indicating murmur or buzz, and “voice” indicating voiced.
d ”,“ compac ”that represents the concentration of the spectrum in the middle frequency band
t "and" diffuse "representing the dispersion of the spectrum in the middle and high frequency bands, etc. In addition, the power of the input waveform which is the acoustic parameter is used.

The outline of the above method of generating a narrower value from the discrimination feature value is as follows.

Now, the value of the narrowing is represented by numerical values of 0, 1, and 2, 0 indicates a closed state (hereinafter referred to as C = 0), and 1 indicates a state where the narrowing is formed somewhere in the vocal tract ( Hereinafter, it is assumed that C = 1, and 2 indicates an open state (hereinafter, C = 2). Then, the relationship between each narrow value and the appearance of each discrimination feature value is as follows.

When C = 0, if voiced, the values of the discrimination features “murmur / buzz” and “nasal” increase. Further, if there is no voice, all the discrimination feature values become small.

When C = 1, the values of the discrimination features “strident”, “diffuse” and “cmpact” increase.

When C = 2, the value of the discrimination feature “voiced” increases. Also, the power is increased.

Therefore, a narrower value can be generated by regularizing such a relationship and applying this rule to each discrimination feature value obtained for each frame of the input speech (the first method). Further, a neural network as described in detail below is used, and a discrimination feature value having a known relationship between the discrimination feature value and the narrow value is input to the neural network as described above. Learning is performed so as to approach the target narrow value. Then, a narrower value can be generated using the neural network for which the learning has been completed (the second method).
Furthermore, when learning the above neural network,
If the acoustic parameters extracted by the feature extraction unit 3 are used as inputs, a narrower value can be directly generated from the acoustic parameters using a neural network (the third method).

Hereinafter, the above-mentioned three techniques will be sequentially and specifically described.

First Embodiment This embodiment relates to a method for obtaining a narrower value from the above-mentioned discrimination feature value in accordance with a rule created based on knowledge.

In this method, the relationship between the narrow value and the discrimination feature value as described above is regularized by a rule including a condition part and an execution part, and stored in a rule file of the narrow value generation part 5. Then, the discrimination feature extracting unit 4 obtains each discrimination feature value for each frame of the input speech, and applies the obtained discrimination feature value to the condition part of the rule to select a rule that satisfies. Then, by executing the execution unit of the selected rule, a narrower value of the current frame is generated.

At that time, the narrower value may not be determined only by the discrimination feature value of one frame. Therefore, the condition of the condition part of the above rule is that the discrimination feature in the previous frame, the discrimination feature in the current frame, The value is determined using knowledge taking into account the value, the discrimination feature value in several frames before and after, the power value of the input speech waveform, and the increase / decrease value thereof.

Next, an outline of the knowledge used when determining the condition of the condition part of each rule will be described.

One value CI- ₁ of narrowing of the previous frame from the combination of the values Ci of the narrowing of the current frame, the nine types shown below when classifying knowledge.

(A) In the case of Ci- ₁ = 0 and Ci = 0 This is the case where the closure in the vocal tract is continued, and generally there is little change in each discrimination feature, and in the case of voice, the discrimination feature "murm"
ur / buzz "and" nasal "are large, and when voiced to unvoiced, discriminating features" murmur / buzz "," voiced "
And the value of "power" decreases. Conversely, when changing from unvoiced to voiced, the values of the discrimination features “murmur / buzz”, “voiced” and “power” increase.

(B) In the case of Ci- ₁ = 0 and Ci = 1 This is the case where a slight gap is created from closure in the vocal tract.
In the case of voiced, the discrimination features “nasal”, “strident”, “diffu
The value of one or more of the values of “se” and “compact” increases, the value of the discrimination feature “murmur / buzz” decreases, and the discrimination features “diffuse”, “strid”
One or more of the values of "ent" and "compact" increases, and power increases in the case of unvoiced.

(C) Case of Ci- ₁ = 0 and Ci = 2 This is a case where the closure in the vocal tract is suddenly opened, accompanied by an increase in power. When it appears at the beginning of a word, it changes from unvoiced to voiced, and the value of the discrimination feature “voiced” increases. In the case of voice, the value of the discrimination feature “murmur / buzz” decreases.
Near the boundary, the value of the discrimination feature “nasal” increases.

(D) Case where Ci- ₁ = 1 and Ci = 0 This is a case where a gap in the vocal tract is closed, accompanied by a decrease in power. At least one of the values of the discrimination features “strident” and “compact” decreases. The value of the discrimination feature "diffuse" may increase near the boundary, but the value immediately decreases. Discriminating feature “murmur / buzz” if voiced
Increases.

(E) Case of Ci- ₁ = 1 and Ci = 1 This is the case where the gap in the vocal tract continues. Discrimination feature “stri
At least one of the values of "dent", "diffuse" and "compact" is large.

(F) In the case of Ci- ₁ = 1 and Ci = 2 This is a case where a gap in the vocal tract is opened, accompanied by an increase in power. Discriminating features “strident”, “diffuse” and “comp
At least one of the values of "act" is decreased.

(G) In the case of Ci- ₁ = 2 and Ci = 0 This is a case in which a closure occurs in any of the vocal tracts, accompanied by a decrease in power. In the case of silence, the values of all discrimination features decrease. In the case of voice, the value of the discrimination feature “murmur / buzz” increases, and the discrimination features “strident”, “diffuse”, and “compact”
Decreases.

(H) In the case of Ci- ₁ = 2 and Ci = 1 This is the case where a gap is formed in any part of the vocal tract, with a decrease in power. Discrimination features “strident”, “diff
At least one of the values of "use" and "compact" increases.

(I) Case of Ci- ₁ = 2 and Ci = 2 This is the case where the opening in the vocal tract continues. Discrimination feature “stri
The values for "dent" and "murmur / buzz" do not increase sharply.

When making the above rules using the above knowledge, as shown in FIG. 4, based on the spectrum and waveform of the voice waveform 21 in advance and the label 22 obtained based on the spectrum and waveform, as shown in FIG. It is made to satisfy the narrow value 23 determined by the above. Then, they are arranged in an optimal order and stored in the rule file. At this time, as the data of the speech waveform for making the rule as described above, the speech waveform 21 obtained from the database consisting of many words generated by many speakers is used.

Next, an example of the above-mentioned rule used when calculating a narrower value of the current frame created based on the above knowledge will be described.

Rule (1): if ((Ci- ₁ = 0) · (Δpower> 0) · (strident> Const1)) then (Ci ← 1) Rule (2): if ((Ci− ₁ = 1) · (Δstrident > 0)) then (Ci ← 1) Rule (3): if ((Ci− ₁ = 1) · (strident> Const2)) then (Ci ← 1) Rule (4): if ((Ci− ₁ = 1) ) · (Δpower> 0) · (strident <Const3)) then (Ci ← 2) where, Const1: constant Const2: constant Const3: constant i: frame number of current frame i− ₁ : frame number of previous frame Ci : Narrower value of the current frame Ci- ₁ : Narrower value of the frame before one frame Also, Δ represents an increase or decrease of the discrimination feature, and is expressed by the following equation.

ΔF = (F (i) + F (i + 1) + F (i + 2)) − (F (i−1) + F (i−2) + F (i−3)) where i: frame number of the current frame F: discrimination Feature values (for example, “strident” and “power”
F (i): discriminating feature value of i-frame The meaning of the above rules is as follows.

Rule (1): a constant Const in which the narrow value before one frame is 0, the power is increased, and the value of the discrimination feature “strident” is
If it is greater than one (eg, the beginning of a fricative), the narrow value of the current frame is set to one. However, Const1
Is about 1/4 of the maximum value of the discrimination feature "strident".

Rule (2): If the narrowing value of the previous frame is 1 and the value of the discrimination feature “strident” is increasing (for example, fricatives continue), the narrowing value of the current frame is changed to Set to 1.

Rule (3): If the narrowing value of the previous frame is 1 and the value of the discrimination feature “strident” is larger than a certain constant Const2 (for example, the fricative sound continues), the narrowing value of the current frame is used. To 1. However, the setting value of Const2 is
About 1/2.

Rule (4): a constant Const in which the narrow value before one frame is 1, the power increases, and the value of the discrimination feature “strident” is
If less than 3 (for example, at the beginning of a vowel)
Set the narrow value of the current frame to 2. However, the setting value of Const3 is about 1/2 of Const2.

FIG. 2 shows a flowchart for generating a narrower value. Hereinafter, the narrow value generation processing will be described in detail with reference to FIG.

In step S1, a discrimination feature value is obtained for each frame by the discrimination feature extraction unit 4 from the feature amount extracted by the feature extraction unit 3.

In step S2, one rule is extracted from a plurality of rules stored in the rule file.

In step S3, it is calculated from the discrimination feature value F extracted in step S1 and the discrimination feature values in the frames Ci- ₁ , Const1, Const2, and Const3 stored in the storage unit of the narrow value generation unit 5 and each frame. Using the obtained ΔF, it is determined whether or not the condition part (if part) of the extracted rule is satisfied. As a result, if the condition part is satisfied, step S4
Otherwise, return to step S2.

In step S4, the rule execution part (then
) Is performed to generate a narrower value Ci of the current frame.

In step S5, it is determined whether the current frame is the last frame. If the result is not the last frame, step S1
To generate a narrower value for the next frame; otherwise, end the narrower value generation process.

In a certain narrow state, transition of the narrow state when each of the above rules is applied is as shown in FIG. That is, if the narrow value of an input voice in an arbitrary frame is C = 0, and if the rule (1) is applied in generating the narrow value in the next frame, the execution unit of the rule (1) Is executed, and a narrower value C of the next frame is obtained.
= 1 is generated, and transition is made as shown by the arrow (A). Less than,
Similarly, if the above rule (2) or rule (3) is applied in the state of C = 1, a narrower value C = 1 for the next frame is generated, and the transition is made as shown by the arrow (B). Further, if the above rule (4) is applied in the state of C = 1, a narrower value C = 2 of the next frame is generated, and the transition is made as shown by the arrow (C).

As described above, according to the present embodiment, since a narrow value is generated based on the discrimination feature value which is a feature amount related to human articulation, it is possible to generate a narrow value which is not affected by the frequency variation by the speaker. Can be.

The narrow information for each frame of the input voice generated as described above is grouped into frame units having the same narrow value, and can be used for restricting a matching path at the time of matching in voice recognition. . In addition, the obtained narrower value of each frame of the input voice is matched with the narrower value of the standard pattern, so that the narrower value is used for calculating the distance between the characteristic pattern of the input voice and the standard pattern. You can also.

Second Embodiment This embodiment relates to a method for automatically obtaining a rule for generating a narrow value by the narrow value generating unit 5 shown in FIG. 1 using the above neural network.

Here, an outline of the neural network will be described. Neural networks are, for example, "" Introduction to Computing
With Neural Nets ", RP Lippman, IEEE AS
SP Magazine, Nikkei Electronics August 10, 1987
As described in No. 427, this is a network that mimics the structure of the human brain, in which a plurality of units corresponding to brain neurons are connected in a complex manner.

The structure of the unit includes a portion that receives an input from another unit, a portion that converts the input according to a certain rule, and a portion that outputs the converted result. The plurality of units form a hierarchical network composed of an input layer, an intermediate layer, and an output layer, as will be described in detail later. Coupling coefficients indicating the strength of coupling are attached to coupling portions with other units. ing.

The coupling coefficient indicates the strength of coupling between the units, and changing the value of the coupling coefficient changes the network structure. That is, the neural
Network learning means that data belonging to one of two events having a certain known relationship is sequentially input to an input layer of the network formed in the hierarchical structure, and then output to an output layer. The coupling coefficient is changed so as to reduce a difference between output data and data (target value) belonging to another event corresponding to the input data. In other words, the structure of the network is changed so that when data belonging to one of two events having a predetermined relationship is input, data corresponding to the other event is output.

The neural network used in this embodiment has a structure as shown in FIG. In other words, this neural network consists of input layers 3 in order from the bottom in the figure.
1. It has a three-layer structure consisting of an intermediate layer 32 and an output layer 33.
The input layer 31 includes 35 units 34, 34,.
Are provided with seven units 35, 35,..., And the output layer 33 is provided with three units 36, 37, 38. Here, the unit 36 of the output layer 33 outputs a narrow value “C = 0”, the unit 37 outputs a narrow value “C = 1”, and the unit 38 outputs a narrow value “C = 1”.
= 2 ". Each unit 34,34,
Are connected to all units 35,..., 35 of the intermediate layer 32, respectively. The units 35, 35,... Of the intermediate layer 32 are output layers, respectively.
33 are connected to all units 36, ..., 36. However, the units in each layer are not connected.

The neural network having the above structure is stored in the storage unit of the narrowing value generator 5 together with the coupling coefficient.

Next, according to the relationship between the discrimination feature value and the narrow value as described above, learning of the neural network having the above configuration is performed as follows using an error back propagation algorithm.

First, input data is input to each unit 34, 34,... Of the input layer 31. Then, this input data is
, Are converted using a predetermined conversion formula (generally, a threshold function or a sigmoid function) and transmitted to the units 35, 35,. At this time, the sum of values obtained by multiplying the output values of the units 34, 34,... Of the input layer 31 by the above coupling coefficient is input to each unit 35, 35,. Similarly, each unit 35, 3 of the intermediate layer 32
5,... Convert the values input from the input layer 31 by a predetermined conversion formula, multiply the values by the above coupling coefficient, and
Output to 36,36,36. Further, similarly, each unit 36, 37, 38 of the output layer 33 converts the value input from the intermediate layer 32 by a predetermined conversion formula to obtain a final output value.

Next, this output value is compared with a desired output value (target value) for the input value, and the value of the coupling strength (coupling coefficient) between the units is changed so as to reduce the difference. That is, the amount of change in the coupling coefficient between the units is calculated from the difference between the output value from the unit when a certain input value is input to each unit from another unit and the target value for the input value. It is.

In this embodiment, each unit 34, 34,.
Includes the discrimination feature values for a total of five frames of the current frame and two frames before and after the current frame (the discrimination features for one frame are “strident”, “nasal”, “murmur / buzz”, “voiced”,
, "Diffuse", "compact", and "power"), that is, the 35 differentiating feature values are input to the units 34, 34,...

At the time of learning, a narrower target value is C = 0 or C =
Only the unit corresponding to 1 or C = 2, that is, the unit corresponding to the target narrow value Ci (for example, the unit 36 if C = 0) outputs “1”, and the other unit (for example,
If C = 0, the units 37 and 38) output the above 35 kinds of discrimination feature values that output “0” to each of the units 34, 34,
Enter in ... The units 36, 37, and 38 of the output layer 33 include units corresponding to the target narrow value Ci (for example, C
If = 0, "1" is input to only the unit 36, and "0" is input to other units (for example, units 37 and 38 when C = 0). Then, the amount of change in the coupling coefficient with respect to the target value is determined by the error back propagation algorithm, and a new coupling coefficient between the units is set.

In this case, as the target values to be input to the units 36, 37, and 38 of the output layer 33, as shown in FIG. This is obtained by obtaining a label 22 based on the audio waveform 21 in advance, and converting the label 22 into a narrow value using a relation table between the narrow value and the label as shown in Table 1.

When the neural network is used to obtain a narrower value for the actual input speech, each of the units 34, 34,... Enter the feature value. Then, the output values of the units 36, 37, and 38 of the output layer 33 are sequentially obtained in accordance with the calculated coupling coefficient between the units, and the unit that outputs the largest value (one of the units 36, 37, and 38) ) Is used as the narrow value of the current frame.

In the present embodiment, in the neural network in which sufficient learning has been performed as described above, the structure of the network corresponds to a set of the 35 discrimination feature values when 35 discrimination feature values are input. It has been changed to output the correct narrow value Ci. In other words, a rule for generating a narrower value is automatically generated. Therefore, the neural network after learning can output a correct narrow value Ci for all input values.

As described above, according to the present embodiment, since a narrow value is generated based on the discrimination feature value which is a feature amount related to human articulation, it is possible to generate a narrow value which is not affected by the frequency variation by the speaker. Can be.

Third Embodiment In the present embodiment, a rule for generating a narrow value is automatically created by the narrow value generating unit 5 in FIG. 1 using the above-described neural network, and the narrow value is directly obtained from the acoustic parameters. The method of generating.

In this embodiment, in order to generate a narrower value directly from the acoustic parameter as described above, the discrimination feature extracting unit 4 in FIG. 1 is removed, and the acoustic parameter extracted by the feature extracting unit 3 is directly narrowed. The value is input to the value generator 5.

The neural network used in this embodiment has a structure similar to that of FIG. However, in this embodiment, the input layer 31 of the neural network has 80 units 34, 34,..., And the intermediate layer 32 has 10 units.
35, 35, ... are arranged, and the output layer 33 has three units 36, 37, 38
Is arranged. Here, as in the second embodiment, the output layer
The unit 36 of 33 outputs a narrower value “C = 0” and the unit 36
37 outputs a narrower value "C = 1" and unit 38 outputs a narrower value "C = 2".

The neural network having the above structure is stored in the storage unit of the narrowing value generator 5 together with the coupling coefficient.

In the case of the present embodiment, each unit 34, 34,.
Is the value of the output power of each of the above BPFs (output from 16 channel BPFs per frame) as acoustic parameters for a total of 5 frames of the current frame and 2 frames before and after the current frame, that is, the output of all 80 BPFs Enter the values in the input layer 31
Are input to the units 34, 34,.

At the time of learning, in the same manner as in the second embodiment, a unit in which the narrower target value is C = 0, C = 1, or C = 2, that is, a unit corresponding to the narrower target value Ci (For example, if C = 0, only the unit 36) outputs “1”, and the other units (for example, if C = 0, the unit 36)
37, 38) input the output power values of the 80 BPFs that output "0" to the units 34, 34,... Of the input layer 31. The units 36, 37, and 38 in the output layer 33 have narrower target values.
"1" is input only to the unit corresponding to Ci (for example, unit 36 if C = 0), and other units (for example, C =
If it is 0, "0" is input to the units 37 and 38). Then, the amount of change in the coupling coefficient with respect to the target value is determined by the error back propagation algorithm, and a new coupling coefficient between the units is set. Units 36, 37 of output layer 33,
The target value to be input to 38 is obtained from the label 22 and the above-mentioned Table 1 which are visually determined in advance based on the audio waveform 21 as shown in FIG.

When the neural network is used to obtain a narrower value for the actual input speech, the units 34, 34,... Enter the output value. And the output layer 3
The output values of the units 36, 37, and 38 are sequentially obtained in accordance with the calculated coupling coefficient between the units, and the narrower value Ci corresponding to the unit outputting the largest value is the narrower value of the current frame. That is.

In this embodiment, in the neural network on which sufficient learning has been performed as described above, a rule for directly generating a narrower value from the output value (acoustic parameter) of each BPF is automatically generated. Therefore, the neural network after learning can output a correct narrow value Ci for all input values. At that time, learning of the neural network
A structure in which the internal structure of the network once generates a feature value equivalent to the discrimination feature value from the input acoustic parameters (BPF output value) and converts the feature value corresponding to the discrimination feature value to a narrower value It is considered that it has been changed.

As described above, according to the present embodiment, the feature extracting unit 3 is used.
Since a narrower value is generated from the output value of the BPF group (that is, the acoustic parameter) via the value of the discrimination feature, which is a feature quantity related to human articulation, the narrower value is not affected by the speaker's frequency variation from the acoustic parameter. Can be generated directly.

The types of discrimination features used in the first and second embodiments are not limited to the above seven types.

In the second and third embodiments, the number of units of the input layer 31 and the intermediate layer 32 constituting the neural network is appropriately changed according to the number of types of discrimination features and the number of frames to be inputted. Needless to say, this is acceptable.

<Effects of the Invention> As is apparent from the above description, the speech feature extraction device according to the first aspect of the present invention includes the discrimination feature extraction unit and the narrowing value generation unit, and is based on the frequency components obtained from the input speech. Therefore, a discrimination feature value representing the degree of discrimination characteristics related to human articulation was obtained, and the narrowed value was generated from the obtained discrimination feature value according to a predetermined rule.
From the discrimination feature value, it is possible to generate a narrow value in the vocal tract that does not depend on the variation of the frequency characteristic by the speaker.

Further, the rule in the speech feature extraction device of the invention according to claim 2 is a rule for generating a narrow value of the current frame based on the discrimination feature values of the previous frame and the current frame. The above narrow value can be generated.

The speech feature extraction device according to the third aspect of the present invention further includes a discriminating feature extraction unit and a narrowing value generation unit having a learning function, and based on a frequency component obtained from the input speech, adjusts human articulation. A discrimination feature value indicating the degree of the related discrimination feature is obtained, and a rule for generating the narrower value from the discrimination feature value is created by learning using an error back propagation algorithm, and the discrimination is performed according to the rule created by the user. Since the narrower value is generated from the feature value, a narrower value in the vocal tract that does not depend on the variation of the frequency characteristic by the speaker can be generated from the discrimination feature value.

Further, the speech feature extraction device according to the fourth aspect of the present invention is provided with a narrowing value generation unit having a learning function, and the feature extraction unit extracts the narrowing value from the acoustic parameters extracted when obtaining a frequency component from the input speech. A rule for generating a value is created by learning using an error back propagation algorithm, and a narrower value is generated from the acoustic parameter according to the rule created by the user. Narrow values in the vocal tract that do not depend on the variation of the vocal tract can be directly generated.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention, FIG. 2 is a flowchart of generating a narrow value in the first embodiment, and FIG. 3 is a narrow state in which each rule in the first embodiment is applied. FIG. 4 is an explanatory diagram of rule creation in the first embodiment and target value creation in learning in the second and third embodiments, and FIG. 5 is a neural network in the second and third embodiments. FIG. 3 is an explanatory diagram of the structure of FIG. 1 microphone 2 amplifier 3 feature extracting unit 4 discriminating feature extracting unit 5 narrowing value generating unit 6
… Word recognition unit, 7… feature pattern storage unit, 8… result display unit, 21… voice waveform, 22… label, 23… narrower value, 31… input layer, 32… middle layer, 33 ... Output layer, 34 ...
... Input layer unit, 35 ... Intermediate layer unit, 36 ...
A unit for outputting C = 0 in the output layer, 37... C in the output layer
= 1 unit, 38 ... Unit for outputting C = 2 in the output layer.

Continuation of front page (56) References JP-A-59-139100 (JP, A) JP-A-63-201698 (JP, A) JP-A-2-32400 (JP, A) JP-A-2-35500 (JP, A) , A) JP-A-64-33598 (JP, A) IEICE Technical Report SP 87-137, P53-60 March 1988 Shimizu "Sound Articulation and Perception" (1983-58-15) Shinozaki Shorin p. 58-77 (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 7/08 G10L 9/10 301 G10L 3/00 539 JICST scientific and technical literature file

Claims (4)

(57) [Claims] An audio feature extracting apparatus for analyzing the frequency of an input voice and extracting a feature of the voice from the obtained frequency component, wherein the voice feature extracting apparatus relates to human articulation based on the frequency component obtained by the frequency analysis. A discrimination feature extraction unit that determines a discrimination feature value indicating the degree of the discrimination feature, and a rule that generates a narrow value indicating the degree of narrowing of the vocal tract, which is one of the feature values, from the discrimination feature value. A narrowing value generating unit that generates the narrowing value from the discriminating feature value obtained by the discriminating feature extracting unit; and a narrowing value in a vocal tract that does not depend on a change in a frequency component by a speaker from the discriminating feature value. An audio feature extraction apparatus characterized by generating a speech. 2. A speech feature extracting apparatus according to claim 1, wherein said rule for generating a narrow value generates a narrow value for a current frame based on at least a discrimination feature value of a previous frame and a current frame. An audio feature extraction device characterized by rules. 3. A speech feature extraction device for performing frequency analysis of an input speech and extracting a speech feature amount from the obtained frequency component, wherein the speech feature extraction device relates to human articulation based on the frequency component obtained by the frequency analysis. A discrimination feature extraction unit that obtains a discrimination feature value representing the degree of the discrimination feature that has been created, and a rule that generates the narrower value from the discrimination feature value is created by learning using an error backpropagation algorithm.
According to the created rule, a narrowing value generation unit that generates the narrowing value from the discrimination feature value obtained by the discrimination feature extraction unit is provided. From the discrimination feature value, the narrowing value generation unit does not depend on a change in a frequency component by a speaker. An audio feature extraction device for generating a narrow value in a vocal tract. 4. A speech feature extraction device for analyzing a frequency of an input speech by a feature extraction unit and extracting a speech feature amount from the obtained frequency component, wherein the feature extraction unit extracts the frequency component when obtaining the frequency component. A rule for generating the narrow value from the acoustic parameter is created by learning using an error back propagation algorithm, and according to the created rule, a narrow value generating unit for generating the narrow value from the acoustic parameter is provided. A speech feature extraction apparatus characterized by generating a narrow value in a vocal tract that does not depend on a change in a frequency component by a speaker from the acoustic parameters.