JPH06324695A - Voice recognizer - Google Patents

Abstract

(57) [Abstract] [Purpose] To realize a voice recognition device that can handle unspecified speakers. A voice recognition device comprising a feature extraction unit, a plurality of feature conversion units, a matching determination unit paired with the feature conversion unit, an input generation unit, and a recognition unit. [Effect] It is possible to recognize the voice of an unspecified speaker without requiring speaker adaptation processing that requires an extremely large amount of processing. As a result, the voice recognition device can be made extremely small.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, and more particularly to a voice recognition device for recognizing a voice of an unspecified speaker.

[0002]

2. Description of the Related Art A speech recognition apparatus is roughly divided into two parts. One is a feature extraction unit and the other is a recognition unit. Each of these has challenges and they are interrelated.

Considering the recognition unit, the recognition means practically used in the conventional speech recognition apparatus are the dynamic programming (DP) method, the hidden Markov model (HMM) method, and the back-propagation learning method. There is a method using a multilayer perceptron type neural network (MLP method). Details of these are described in, for example, Seiichi Nakagawa "Speech Recognition by Probabilistic Model" (IEICE), and Nakagawa, Kano, and Higashikura "Speech / Hearing and Neural Network Model" (Ohmsha). There is.

A problem common to the DP method and the HMM method is that the data to be a teacher and the data to be recognized require a start end and an end. In order to perform processing that does not seem to depend on the starting end point, there is a method of performing processing for all possible starting end points and finding the starting end point that gives the best result by trial and error.
However, considering the case of detecting a portion of data belonging to a certain category from a pattern of length N, there is a possibility of the order of N as the start end, and there is a possibility of N at the end. There is a possibility of order. In other words, there is a possibility that there will be N surplus orders as a combination of the start end and the end. Therefore, in this case, the recognition process must be performed for all of these very large combinations. Therefore, there is a problem that the processing takes a very long time.

Further, before the quantitative problem of the number of combinations, there is a more essential qualitative problem in the assumption itself of the existence of the beginning and end. If the condition that the input data is composed of only one data of a certain category, the start end is obvious. However, when data of one or more categories are continuous, the boundary is not obvious. Rather, in time-series information such as voice, such a boundary does not exist clearly, and it is considered that data in two consecutive categories changes from one to the other through a transition region in which the information overlaps. Therefore, assuming the beginning and end of data has a very big problem in its accuracy.

In the case of the MLP method which is another method of the conventional method, it is not necessary to particularly assume the beginning and end of such data. However, in place of that, a new start / end problem of the range of input data occurs. In other words, the MLP method is basically a method for recognizing static data, and when recognizing time-series data using it, by inputting data in a certain time range as one input data , There is a problem that time information must be processed equivalently. This time range must be fixed due to the structure of the MLP.

On the other hand, the length of the time-series data often varies greatly depending on the category and even within the same category. For example, taking phonemes in speech processing as an example, the average lengths of vowels, which are long phonemes, and plosives, which are short phonemes, differ by a factor of 10 or more. Also in the same phoneme category,
The length in the actual voice fluctuates more than twice. Therefore, if the input range of data is set to the average length of all phonemes, when recognizing a short phoneme, a large number of data other than the recognition target will be included in the input data, and vice versa. When recognizing a long phoneme, only a part of the data to be recognized is included in the input data. All of these are factors that reduce cognitive ability. In addition, even if you set a different length for the data input range for each phoneme,
The problem is similar because the length within the phoneme category varies. Such a problem is commonly found in time series information.

In addition to the problems described above, an important issue in constructing a practical voice recognition device is how to accurately recognize the voices of many speakers having various characteristics. It means to recognize it.

The main conventional methods to deal with this are:
This is a method in which a recognition unit corresponding to one standard feature amount is prepared, and the speaker-dependent features extracted by the feature extraction unit are converted into features corresponding to the standard speaker and used. Such a method is called speaker adaptation. FIG. 3 schematically shows this. In the figure, 301 is a feature extraction unit,
Reference numeral 302 is a feature conversion unit, 303 is a recognition unit, and 304 is an output unit. Specifically, as an example of speaker adaptation processing, that is, a method of configuring a feature conversion unit, Furui, "Speaker adaptation by hierarchical clustering of spectrum space" (The Institute of Electronics, Information and Communication Engineers, Speech Study Group material SP88-21) Etc. However, a large amount of adaptive data and a large amount of calculation for the processing are required to actually execute such adaptive processing and obtain good results.

Another method for dealing with a large number of speakers is
Speakers with similar characteristics are classified into several groups, a recognition means is prepared for each of the groups, and voice recognition is performed by a recognition means for a speaker group similar to the characteristics of the speaker to be recognized. Is to say. FIG. 4 schematically shows it. In the figure, reference numeral 401 is a feature extraction unit, 402 is a plurality of recognition units, 403 is an output selection unit, and 404 is an output. In this method, at least in the final recognition device, a large amount of adaptation data such as speaker adaptation processing is not required,
In addition, the adaptive processing itself is unnecessary.

However, this method has a problem that the system becomes large because it has a plurality of recognition means, and it is difficult to select an optimum recognition means.

[0012]

As described above, the conventional DP method and HMM method require the beginning and end of data to be handled. Apparently, in order to relax this restriction, it is necessary to process all the combinations of the start end and the end. However, there is a problem that this requires a great deal of processing.

In addition, the MLP method requires the beginning and end of the input range during learning. However, a problem with any of these methods is that in general time-series information, the beginning and end of data cannot be defined in principle. Moreover, there is a problem that forcibly assuming it lowers cognitive ability.

Further, there is a problem that a large amount of adaptation data and its processing are required for speaker adaptation for recognizing data of an unspecified speaker. Further, the speaker group selection has a problem that the system becomes large and it is difficult to realize the selection means.

[0015]

According to the present invention, 1) a feature extraction unit, a plurality of feature conversion units that receive the extracted features, a matching determination unit paired with the feature conversion unit, and a feature A voice recognition device, comprising: an input generation unit that receives the output of the conversion unit and the matching determination unit; and a recognition unit that receives the generated input as an input, 2), and the recognition unit is a neural network. , Each nerve cell-like element that constitutes the neural network is internally based on the internal state value storage means, the internal state value stored in the internal state value storage means, and the input value input to the nerve cell-like element. The speech recognition apparatus according to 1), further comprising an internal state value updating means for updating the state value, and an output value generating means for converting the internal state value by the internal state value storage means into an external output value. 3), the said conformity The determination unit includes at least two feature input units, a feature mapping unit that maps and transforms one of the input features, and a comparison unit that compares the other input feature and the mapped feature. The speech recognition device according to 1) or 2) above, 4), wherein the feature mapping unit is a neural network, and each neuron-like element that constitutes the neural network has an internal state value. Storage means, internal state value updating means for updating the internal state value based on the internal state value stored in the internal state value storage means and the input value input to the nerve cell-like element, and the internal state value storage means 5. The voice recognition device according to any one of 1) to 3) above, further comprising: an output value generation unit that converts the internal state value into an external output value, 5), and the internal state value updating unit. ,Previous The internal state value and the input value each include a weighted summing unit that sums and accumulates weights, and the internal state value storage unit includes an integrating unit that integrates the values accumulated by the weighted summing unit. The output value generating means includes an output value limiting means for converting the value obtained by the integrating means into a value between an upper limit value and a lower limit value set in advance, 1).
To 6), 6), Xi is the internal state value of the i-th neuron-like element that constitutes the neural network, and n weighted external input values to that element Is Zj (j is 0 to n, n is a non-negative integer), and the time constant is τi, the internal state value updating means

[0016]

[Equation 2]

The voice recognition device according to any one of 1) to 5) above, wherein the internal state value Xi is updated to a value satisfying the above condition. 7) The weighted external input value Zj is at least , The output of the i-th neuron-like element itself is weighted, the output of the other neuron-like elements that make up the neural network is weighted, and is given from outside the neural network The voice recognition device according to any one of 1) to 6) above, wherein the input value includes one or more of a weighted value added to a fixed value, 8), and the input generation. Section selects any one of the outputs of a plurality of feature conversion sections input to the speech recognition apparatus according to any one of 1) to 7) above, 9), the input generation section Type in it A speech recognition apparatus according to any one of 7) from the 1), characterized in that to produce a linear combination of the outputs of the plurality of feature transformation unit to be.

[0018]

【Example】

(Embodiment 1) FIG. 5 shows the flow of processing of the feature extraction unit. The input speech is digitized, partial data called a frame is discretely cut out, and a feature vector is extracted from the frame by a method such as linear prediction analysis. In the following, the sequence of feature vectors extracted in this way will be described as the output of the feature extraction unit.

FIG. 1 is a schematic diagram of the configuration of the speech recognition apparatus of the present invention. In the figure, reference numeral 101 is the feature extraction unit as described above, 102 is the feature conversion unit, 103 is the matching determination unit, 104 is the input generation unit, and 105 is the recognition unit.

First, the operation of the recognition unit of the present invention will be described. The recognition unit of the present invention is composed of a neural network. This neural network is different from the conventional MLP method in that a nerve cell-like element (hereinafter simply referred to as a node) constituting the neural network has a function as shown in FIG. In the figure, reference numeral 601 indicates internal state storage means, 602 indicates internal state update means, 603 indicates output value generation means, and 604 indicates the entire node.

Here, the internal state value of the i-th node is X
i, Zj (j is an integer from 0 to n, n is a non-negative integer) input to the weighted input value, and τ is a time constant.
If i, the operation of the above internal state value updating means is expressed by the following equation, for example.

[0022]

[Equation 3]

In such a node whose operation is given by a differential equation, the node itself can have the ability to process time information. A more specific example of the above equation is that the output of the jth node is Yj,
The connection weight from the j-th node to the i-th node is Wi
When the bias value to the j-th and i-th nodes is θi and the net external input value given to the i-th node is Di, it is expressed by the following equation.

[0024]

[Equation 4]

Here, the bias value θi can be taken into Wij as a weighted connection to a node having a fixed output value in calculation. Further, what is given as Di is the component of the feature vector described above in the case of the voice recognition device. Of course, this value is given only to the assigned node of the input node in the neural network.

The relationship between Xi and Yi described above is given by a simple linear function, a threshold function, or a non-linear function. In this example, a positive / negative symmetrical sigmoid function (also called a logistic function) given by the following equation was used.

[0027]

[Equation 5]

Now, in order to give the neural network a desired function, learning processing is required. In the above example, the parameter to be adjusted is the connection weight Wij, and the learning amount for the Wij, that is, the learning process for the node described above, is expressed as follows, for example. There is.

[0029]

[Equation 6]

Here, Ci is an evaluation value of a certain learning, and E
j is an error evaluation value obtained from the difference between the desired output (hereinafter referred to as the teacher output) and the actual output. This error evaluation value is
For example, the teacher output may be Ti and the actual output may be Yi, which can be given by the Kullback-leibler distance as in the following equation.

[0031]

[Equation 7]

The above equation is an example when the output range is from 0 to 1, but when the output range is from -1 to 1 as in the case of this embodiment, the following equation equivalent to the above equation is obtained. Is represented as

[0033]

[Equation 8]

For the system connected with the connection weight Wij as described above, the learning evaluation value formula can be specifically expressed as follows using the Kullback-leibler distance as the error evaluation function. .

[0035]

[Equation 9]

By the processing as described above, the output Y
When j and the learning evaluation value Ci are obtained, the correction amount of the connection weight, that is, the learning rule is expressed by the following equation.

[0037]

[Equation 10]

According to the learning rule as shown here, when the neural network is constructed, arbitrary connection forms including layered connection, complete mutual connection and the like are possible. Further, unlike the conventional MLP method or the like, it is possible that the input node and the output node are the same.

In the example described above, the function shown in FIG. 6 can be expressed as shown in FIG. 70 in the figure
1 is a data input means, 702 is a weighted integration means,
Reference numeral 703 represents an integrating means. These constitute an internal state value updating means and an internal state value storage means. Reference numeral 704 represents an output value limiting means such as a sigmoid function or a threshold function as an output value generating means. These functions can also be expressed as shown in FIG. 8 on hardware using an operational amplifier or the like. In the figure, numeral 801 is a data input means and weighted integrating means, 802 is an integrating means, and 803 is an output value limiting means. However, FIG. 8 is an example, and the configuration is not limited to this. For example, a configuration in which the integrating means and the output value limiting means are exchanged is possible.

FIG. 9 shows the configuration of a simple speech recognition apparatus constructed by using such a node. In this figure, six neural networks which are completely connected to each other asymmetrically and have a self-feedback connection are exemplified as the recognition means.

The learning process is necessary for the neural network to have a certain function, and the basic algorithm of the process is described above. However, in this learning process, what kind of data is to be learned is a problem in a category different from that of the learning algorithm, and is a more important problem in order for the neural network to have a certain desired ability.

In this embodiment, one neural network as shown in FIG. 9 is made to recognize data in one category. Temporarily, the data of the category set for the neural network will be called positive data, and the other data will be called negative data. Using these, as input data for learning, data obtained by chaining two data in those categories was used. Two outputs were set, a positive output and a negative output.

The correspondence between the learning output and the input data is as shown in FIG. In other words, when positive data is input, the output indicated by the solid line (positive output) shows a large value,
The output (negative output) indicated by the broken line has a small value. When the input is a negative output, the opposite is assumed. Then, the transition between them when the data is continuous gives the condition that the value of the data and its time derivative are continuous and is as shown in the figure. At this time, it is assumed that one frame of input data corresponds to one frame of output data, and the data of a plurality of frames are not simultaneously input unlike the conventional MLP method.

There are two reasons for giving the input and output as shown in FIG. One is that since the operation of this neural network is expressed by a first-order differential equation, an arbitrary initial value is required. In the example of the figure, the operation is always started with the origin as the initial value, but in the data after the chaining, the state determined by the previous data becomes the initial value. Therefore, by using various combinations of this chain data, this neural network learns the desired operation for various initial values. The other is related to this transition, that is, even if some data is input, the category is not determined at the start of the input. For example, even if the voice / i / is input, it is uncertain whether it is "Ikioi" / i / or "Iiyoi" / i /. In that sense, the transitions of this output were chosen to be symmetric at the midpoint of the data.

By performing such learning, for example, FIG.
The recognition as shown in 9 is possible. The example of FIG. 19 is an example of embedding positive data, which is recognition symmetry, in a plurality of negative data. The data used for learning was only a chain of two data, but as seen in this example, positive data can be recognized from the data of arbitrary length for learning. The effects of this recognition method can be summarized as follows.

1) It is not necessary to strictly and explicitly give and learn the start and end of data as in the conventional example. Therefore, in the conventional example, the time which is proportional to the surplus of the length of the data was required in the recognition processing, but the processing can be performed in the time proportional to the length.

2) Since only the input data at that moment needs to be processed, the memory may be very small.

3) No correction is required for the resulting input length.

4) As described above, continuous processing can be easily performed.

5) The start end and end of the classified data have little influence on the learning result.

6) Only by learning various combinations of two categories of data, it is possible to obtain recognition ability for an arbitrary number of combinations of categories.

In order to further apply this to the input of an unspecified speaker, the present invention has the feature conversion section and the matching determination section shown in FIG.

The concept of the operation of the feature conversion section will be described with reference to FIG. The ellipse indicated by the number 1001 in the figure indicates the spread of the characteristics of the input speaker, and the ellipse indicated by 1002 is the speaker possessed by the recognition unit or the speaker used for learning of the recognition unit (hereinafter abbreviated as standard speaker). Let us show the spread of the characteristics of. If the spreads match, the voice of the input speaker can be accurately recognized by the recognition means by the standard speaker.
However, as shown in FIG. 10a), the recognition accuracy is degraded unless the spreads of the feature quantities overlap.

In the speaker adaptation processing, as schematically shown in FIGS. 10b) and 10c), the feature amount of the input speaker is converted into the corresponding amount of the standard speaker by some feature conversion operation, and the recognition accuracy is improved. It is an attempt to improve. Although various types of feature conversion processing can be considered, a general method is a method using linear conversion represented by the following equation.

[0055]

[Equation 11]

Here, Uj represents the j-th component of the speech feature vector U with an input speaker, Vi represents the i-th component of the standard speaker feature vector V corresponding to the input feature vector U, and Tij is The ij component of the transformation matrix representing operations such as rotation, reduction, and expansion of the vector is shown, and Gi represents the i-th component of the movement vector G representing the movement operation of the vector. Performing the speaker adaptation process is to determine the above Tij and Gi so as to obtain the optimum result. However, a large amount of adaptive data and its processing are required to do this accurately. In addition, when the adaptation data is small, the result is often worse.

On the other hand, the present invention has a plurality of pre-processed feature conversion operations as shown in FIG. Since the feature conversion unit is fixed, it is not necessary to perform adaptive processing with a large amount of data and a large amount of processing each time. Further, by separating such processing from the recognition unit, processing independent of the function of the recognition unit becomes possible.

The quality of the performance of such a processing system depends on
Which of a plurality of pre-existing feature conversion units should be used to convert the actually input data, or what kind of input should be generated by combining the converted features It is determined by the capabilities of the conformity determination unit and the input generation unit that determine things. Various types of conformity determining units are conceivable, but in the present embodiment, a unit as schematically shown in FIG. 2 is considered. In the figure, 201 and 202 are feature input units, 203 is a feature mapping unit that transforms and maps one of the input features, and 204 is a comparison unit between the transformed features and the input features.

Further, in this embodiment, as shown in FIG. 11, the feature mapping section is exemplified by a neural network similar to the recognition means. In FIG. 11, number 1
101 is an input feature vector sequence, 1102 is a neural network forming a feature mapping unit, and 1103 is a comparison unit. Further, in FIG. 11, 20 of FIG.
The feature input unit 1 shown in FIG.
The input given to 1 is illustrated as a case where the input is data immediately before the data input to the feature input unit 1. That is, FIG.
The input of the comparison unit in the first example is the feature vector converted by the feature mapping unit at a certain time point and the feature vector immediately before that. Considering that the feature mapping unit is configured so that the difference between them becomes small, this means that the data before that is restored based on the feature vector at a certain time. Similarly, assuming that the data given to the feature input unit 2 is data after the data given to the feature input unit 1, this means that data after that is predicted based on data at a certain time. It is a thing. When the two data inputs are the same, it means that a simple mapping of the data is performed.

The ability of such feature vector prediction, simple mapping, and restoration depends on the feature data of the speaker used for learning the ability. Therefore, by such processing, it is possible to determine the degree of conformity, which is whether the input feature vector sequence is close to the learned one. Especially predictions,
The method using restoration is more effective for explicitly learning dynamic features in addition to static features. In addition, as for speech, a feature at a certain point of time has a high influence of a feature that follows it, so that the restoration process may give a more accurate result.

12 and 13 show two speakers MAU and MX.
In this example, M is a learning speaker and a speaker suitability determination unit is configured by data restoration to make a determination. The horizontal axis is time,
In the figure, the solid line shows the output of the matching determination unit with MAU as the learning speaker, and the broken line shows the output of the matching determination unit with MXM as the learning speaker. The vertical axis indicates the absolute value of the difference between the feature vectors by the comparison unit. As is clear from the figure, the input of each speaker causes the matching determination unit learned by each speaker to show a smaller data restoration error than the other. In other words, it indicates that it is better suited.
Moreover, this result is obtained at the same time as the start of data input, that is, it is possible to judge the compatibility with a very small amount of data.

The table below shows an example in which the same processing as the above example is applied when an unknown speaker is included. The numbers in the table show the average restoration error of the data when the word string used for learning is input, in arbitrary units.

[0063]

[Table 1]

Table 2 is the same as Table 1 except that the input voice data is data other than the data used for learning, that is, data unknown to the system is input.

[0065]

[Table 2]

Similar results are obtained in this example as in the example described above. In other words, it is considered that the matching determination unit responds to the characteristics of the speaker, not the information specific to the learned input data. Further, the values in the above two tables show similarity, and it is considered that this value serves as an index of the similarity between speakers.

16 and 17, the recognition unit is trained by the data of the speaker MAU, and the feature conversion unit is MXM.
Of MAU features to MAU features and MAU features to MA
This is an example in which two are prepared for conversion into U features. Hereinafter, these conversion units will be written as MXM-MAU conversion units. In this case, the MAU-MAU converter is an identity conversion.
In this example, the input speaker is MMY, and only one data to be recognized is included in the utterance. FIG.
For comparison, the input data is MXM-MA
In this example, an input is generated by performing U conversion and input to the recognition unit. As is clear from the figure, two recognition outputs are obtained, and erroneous recognition is seen.

On the other hand, according to the present invention, as is clear from the above table, the matching determination unit determines that the speaker MMY is more similar to MAU than MXM. Therefore, the input data is subjected to MAU-MAU conversion (identity conversion in this embodiment) and input to the recognition unit. FIG. 17 shows the result, and the expected recognition result is obtained.

(Second Embodiment) The operation of the input generation section of the first embodiment shown above is such that only the data conversion section having the highest degree of conformity is selected, but other operations should be considered. Is also possible.

Consider, for example, the case shown in FIG. It is assumed that the numbers 1401, 1402, 1403 in the figure indicate the spread of the features of the speaker used for learning of the feature conversion unit prepared in advance. Also, the shaded number 1404 indicates the spread of the characteristics of the input speaker. In this example,
According to the processing of the first embodiment, the feature conversion unit corresponding to the data shown in 1401 is selected. But 1
When the similarity between 401 and the input speaker is not so close, the expected result may not be obtained in this process. FIG. 18 shows a recognition example in such a case, and no data to be recognized is detected.

In the case of the case of FIG.
As shown by the arrow, by combining the respective converted data according to the similarity between the respective feature conversion sections and the input data, the accuracy of the approximation of the input speaker can be improved. This combination can be realized by, for example, a linear combination represented by the following equation.

[0072]

[Equation 12]

Here, Tijk and Gik respectively represent a transformation matrix and a movement vector forming the k-th transformation unit. Further, ρk represents the weight of the feature obtained as a result of the k-th conversion. This weight is obtained, for example, by subtracting a constant from the data restoration error shown in the above table and normalizing the reciprocal of the constant so that the total sum becomes 1. FIG. 19 shows the result, and the expected recognition result is obtained.

[0074]

Summarizing the above, the effects of the present invention are as follows: 1) It is not necessary to strictly and explicitly give and learn the beginning and end of data as in the conventional example. Therefore, in the conventional example, the time which is proportional to the surplus of the length of the data was required in the recognition processing, but the processing can be performed in the time proportional to the length.

2) Since only the input data at that moment needs to be processed, the memory may be very small.

3) No correction is required for the resulting input length.

4) As described above, continuous processing can be easily performed.

5) The start end and end of the classified data have little influence on the learning result.

6) Only by learning various combinations of two categories of data, it is possible to obtain recognition ability for an arbitrary number of combinations of categories.

7) In the final recognition device, speaker adaptation processing with a large amount of processing is unnecessary.

8) It is possible to judge the suitability of the speaker with a very small amount of data.

9) As a result of the above, it is possible to downsize the system corresponding to the unspecified speaker.

There are advantages such as the following.

These effects can be applied not only to voice recognition but also to online character recognition or more general time series data processing.

[Brief description of drawings]

FIG. 1 is a block diagram of elements constituting a voice recognition device of the present invention.

FIG. 2 is a block diagram of elements constituting a matching determination unit which is a part of the voice recognition device of the present invention.

FIG. 3 is a block diagram of elements constituting a conventional voice recognition device.

FIG. 4 is a block diagram of elements constituting a conventional voice recognition device.

FIG. 5 is a schematic diagram showing a flow of voice recognition processing.

FIG. 6 is a block diagram of nodes constituting a neural network according to the present invention.

FIG. 7 is a block diagram of nodes constituting a neural network according to the present invention.

FIG. 8 is a schematic diagram of a hardware configuration of a node configuring a neural network according to the present invention.

FIG. 9 is a schematic diagram of a voice recognition device using a neural network.

FIG. 10 is a schematic diagram showing speaker adaptation processing.

FIG. 11 is a schematic diagram showing a conformity determination process of the present invention.

FIG. 12 is a diagram showing an output example by the conformity determination processing of the present invention.

FIG. 13 is a diagram showing an output example by the conformity determination processing of the present invention.

FIG. 14 is a schematic diagram showing speaker adaptation processing.

FIG. 15 is a diagram showing a learning output example of a recognition unit using the neural network of the present invention.

FIG. 16 is a diagram showing a recognition example of the present invention.

FIG. 17 is a diagram showing a recognition example of the present invention.

FIG. 18 is a diagram showing a recognition example of the present invention.

FIG. 19 is a diagram showing a recognition example of the present invention.

[Explanation of symbols]

 101: Feature extraction unit 102: Feature conversion unit 103: Matching determination unit 104: Input generation unit 105: Recognition unit 201: Feature input unit 1 202: Feature input unit 2 203: Feature mapping unit 204: Comparison unit 301: Feature extraction unit 302: Feature conversion unit 303: Recognition unit 304: Output unit 401: Feature extraction unit 402: Recognition unit 403: Output selection unit 404: Output unit 601: Internal state value storage unit 602: Internal state value updating unit 603: Output value generation Means 701: Data inputting means 702: Weighted integrating means 703: Integrating means 704: Output value limiting means 801: Data input, weighting integrating means 802: Integrating means 803: Output value limiting means 901: Voice feature extraction section 902: Neural Network 903: Recognition result output unit 1001: Input feature 1002: Standard feature 1003: Conversion feature 1101: the time-series feature vector 1102: Mapping Neural Network 1103 Comparison 1401: Standard Features 1402: Standard Features 1403: Standard Features 1404: input feature

Claims (9)

[Claims] 1. A feature extraction unit, a plurality of feature conversion units that receive the extracted features, a matching determination unit that makes a pair with the feature conversion unit, and an output of the feature conversion unit and the matching determination unit as inputs. A voice recognition device, comprising: an input generation unit that performs input and a recognition unit that receives the generated input as an input. 2. The recognition unit is composed of a neural network, and each nerve cell-like element constituting the neural network includes an internal state value storage means, an internal state value stored in the internal state value storage means, and its nerve cell. The internal state value updating means for updating the internal state value based on the input value input to the element, and the output value generating means for converting the internal state value by the internal state value storage means into the external output value. The voice recognition device according to claim 1, wherein the voice recognition device is a voice recognition device. 3. The conformity determination unit includes at least two feature input units, a feature mapping unit that maps and converts one of the input features, and another feature that is input and a feature that is mapped. The voice recognition device according to claim 1 or 2, further comprising: a comparison unit that compares 4. The feature mapping unit comprises a neural network, and each nerve cell-like element forming the neural network includes an internal state value storage means, an internal state value stored in the internal state value storage means, and its nerve. It has an internal state value updating means for updating the internal state value based on the input value input to the cell-like element, and an output value generating means for converting the internal state value by the internal state value storage means to an external output value. The voice recognition device according to any one of claims 1 to 3, characterized in that. 5. The internal state value updating means comprises weighted integrating means for adding weights to each of the internal state value and the input value and integrating the weights, and the internal state value storage means:
The output value generating means includes an integrating means for integrating the value integrated by the weighted integrating means, and the output value generating means converts the value obtained by the integrating means into a value between a preset upper limit value and a lower limit value. 5. The voice recognition device according to claim 1, further comprising an output value limiting unit. 6. The internal state value of the i-th neuron-like element forming the neural network is Xi, and n weighted external input values to the element are Zj (j is 0 to n, n is non-negative). And the time constant is τi, the internal state value updating means is The voice recognition device according to any one of claims 1 to 5, wherein the internal state value Xi is updated to a value that satisfies the above condition. 7. The weighted external input value Zj is at least the output of the i-th nerve cell-like element itself with weights added, and the output of another nerve cell-like element constituting the neural network. The sum of weights, the input value given from outside the neural network,
7. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus includes one or more of a fixed value and a weight added thereto. 8. The voice recognition device according to claim 1, wherein the input generation unit selects any one of outputs of the plurality of feature conversion units input to the input generation unit. . 9. The voice recognition device according to claim 1, wherein the input generation unit generates a linear combination of outputs of a plurality of feature conversion units input to the input generation unit. .