JPH064097A - Speaker recognizing method - Google Patents

Abstract

PURPOSE:To precisely recognize a speaker with a small number of learning data by inputting an internal state value and an external input value, and updating the internal state value with the input and converting it into an external output value. CONSTITUTION:This method is equipped with a speech feature extracting means 501, a neural network 502, an input value storage means 504 which stores the input of the neural network 502, and a data comparing means 503 which compares the stored input with the output. In this case, the input value storage means 504 is stored with only input data which is one frame before and compares the output data of a current frame to recognizes the speaker. Namely, a speech feature series can easily be understood from the mechanism of the generation, but the continuous motion of a speaker and a modulation organs is reflected. Then the individuality of the speaker appears as the feature of the motion, so its feature time series is processed to recognize the individual. The process is therefore performed by evaluating the individuality of the input speech by restoring the data right before the current data.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION The present invention relates to a method of recognizing a speaker.

[0002]

2. Description of the Related Art Methods used for speaker recognition are roughly classified into two.
There are two ways. One method is to use the time series of the feature vector of the utterance obtained from the input speech, and the other method is to use the statistical feature quantity obtained by statistically processing the time series. is there. These methods are described in detail in, for example, Chapter 9 of "Digital Audio Processing" by Toshihiro Furui (Tokai University Press).

However, it is necessary to collect a large amount of data in order to perform highly accurate recognition in the method using the statistical feature amount. However, obtaining such a large amount of data from the speaker is extremely difficult in practice. For example, inviting a speaker to be recognized to speak for several minutes in front of the speaker recognition device greatly limits the application of the speaker recognition device. Further, if a statistic is estimated from a small amount of data in order to relax this limitation, there is a problem that the estimation accuracy deteriorates due to the estimation error.

If a large amount of data can be obtained, it is considered that the statistical method described above is a good choice, but if the amount of data is small,
The method of using the speaker feature time series obtained therefrom often gives good results. This is because, compared to statistical quantities, which are mostly static quantities, the method of processing the time series itself can also utilize the dynamic aspect of the speaker's characteristics. On the contrary, unless a method capable of accurately processing such a dynamic characteristic of the speaker, the speaker cannot be recognized from the speaker feature data. This is also mentioned in the above-mentioned "digital speech processing", but although the method for speaker recognition requires special consideration for that, it is essentially speech. The point is that a method very similar to recognition is needed.

Considering the conventional speech recognition methods, the conventional speech recognition methods are roughly classified into DP matching method (DP method), hidden Markov model method (HMM method), back propagation learning method and neural. Method using multi-layer perceptron which is a network (MLP
Law). 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.

When the DP method is used for speaker recognition, first, standard data for each speaker is collected, and at the time of recognition, the correspondence between the input end and the start end is assumed for each input data. Then, the correspondence of the internal elements is changed by various time normalization functions, and the correspondence relationship that minimizes the difference and the difference between the patterns at that time are set as the distance between the input data and the standard pattern, and the distance is minimized. The speaker represented by the standard pattern is used as the recognition result.

In this case, the assumption of the correspondence between the start end and the end is that the distance between the input pattern and the standard pattern increases in proportion to the length of the pattern. For example,
Even if the speaker recognition is performed using a certain word or a sentence, the speaking speed is different for each person, and even the same speaker changes depending on the situation. Therefore, in order to compare distances that do not depend on the pattern length between standard patterns of different lengths, it is necessary to normalize the distances to the standard pattern or the length of the input data. That is, the correspondence of the distance, that is, the beginning and end of the pattern is essential.

In the HMM method, what replaces the standard pattern of the DP method and represents a speaker is an HMM model composed of a plurality of states and a plurality of transitions. An existence probability is given to each state of the HMM model, and a transition probability and an output probability are given to each transition, and these probability values are determined by learning using learning data. With these learned probability values, the HMM model statistically and stochastically represents one speaker.

In the HMM method, at the time of speaker recognition, for each of the HMM models representing each speaker, it is assumed that the beginning and the end correspond to the input data, as in the DP method. Under the condition that the output data sequence is output, the probability of how close the input data is to each speaker is calculated as the probability of transition from the start state to the end state. Then, the speaker represented by the HMM model which maximizes the probability is recognized as the speaker to which the input data belongs.

Here, the HMM model models the time series data in a time series in a statistical and probabilistic form such as states and transitions. Therefore, at the time of learning, it is necessary to specify a portion close to the start state of the learning input data, a portion close to the end state, an intermediate portion, and the like. For that purpose, it is necessary to accurately give the start and end of the data. If the method of giving the starting point is inaccurate and more data than necessary is given to the portion near the starting point, this will reduce the recognition ability of the model. On the contrary, if the learning data lacks necessary data, the input data including the missing data cannot be accurately recognized. As a result, there is a high possibility of being erroneously recognized.

At the time of recognition, the criterion in the HMM method is the transition probability from the initial state to the final state, and the existence probability in the final state. Since this value is the product of the output probability of each component of the input data string, the transition probability, and the existence probability of a certain state, it decreases monotonously and very rapidly depending on the length of the data. Therefore, for example, even if it is said that the existence probability of the final state is 0.5, whether the value is the value for the data string of length 10 or the value for the data string of length 20. Depending on the weight, the meaning or meaning is completely different. Therefore, in order to make a determination that does not depend on the data length, some kind of processing for correcting the data length is required. This requires the length of the data, that is, the beginning and end of the data.

As described above, in both the DP method and the HMM method, a certain unit of data such as the beginning and end of data is required, and only the result for each unit can be obtained. Can not. If the unit of the data is small, the data can be collected easily, but the variation of the data becomes large, and the recognition accuracy deteriorates. Further, if this unit becomes extremely large, it becomes difficult to collect data, the modeling accuracy in the DP method or the HMM method also deteriorates, and as a result, the recognition accuracy deteriorates. That is, there should be an optimal data unit size, which is not a priori determined. In addition, such a method of performing processing while changing the size, or processing of various possibilities of various starting ends and terminations requires a very long processing time.

An even more important problem is that these methods can basically perform the speaker recognition process only on the data having the same contents as the standard data. This is inevitably the process described above is a simple pattern matching process between input data and standard data.

On the other hand, ML which is another method of the conventional method
In the case of the P method, it is possible to obtain any number of output values at any time. In addition, it is not necessary to assume a unit such as the beginning and end of data. However, the conventional MLP method has a problem of a new start and end in terms of the range of input data, not the start and end of data. That is,
The MLP method is basically a method for recognizing static data, and in order to make it recognize time series data,
The temporal structure of the input data must be reflected in the neural network structure in some way. The method most often used as this method is to input data in a certain time range as one input data and process the time information equivalently. However, this time range must be fixed due to the structure of the MLP.

At this time, it is difficult to recognize a temporal feature that exceeds the input time range, and it is also difficult to recognize a temporal feature that is too small compared to the input time range. Is. That is, unnecessary data is inserted before and after the temporal feature to be recognized. On the other hand, the length of the speaker feature time-series data that is input can vary greatly depending on the speaker and even within the same speaker, so such an input range mismatch is very likely to occur. It can happen.

As an example which does not have such a fixed input range, there is a modification of the conventional MLP method when the output is fed back to the input side. An example of this is the case of character recognition. For example, the three-layer BP model with feedback coupling is used, from page 1556 to page 1564 of J74 volume (1991) of the Institute of Electronics, Information and Communication Engineers D-II. "Printed handwritten character recognition".

However, as is clear from the above-mentioned literature, it is difficult for these methods to converge the learning of the neural network, and the learning output (teaching signal) for that purpose is created by trial and error. There is a problem that it must be.

[0018]

As described above, in the conventional speaker recognition method, 1), in the method using the statistical quantity, the data collection is very difficult. It is difficult, and the method of estimating it from a small amount of data is prone to error.

2) The characteristic time series data is converted to the DP method or the HMM.
The method of processing by the method requires processing data of an appropriate length, its start end and its end, and takes a long processing time. Also, it is difficult to obtain results continuously.

3) Also, in this speaker recognition processing, there is a constraint that the input data must have the same utterance content as the standard data.

4) The method of processing the characteristic time series by the conventional MLP method requires the beginning and the end of the input range, and it is difficult to deal with the change in the data length. Also, it is difficult to converge learning.

There are problems such as

[0023]

The speaker recognition method of the present invention for solving the above-mentioned problems is a speaker recognition method using a neural network, wherein the neural network has at least an internal state value storage means, It is configured by a nerve cell-like element having an internal state value updating means for updating the internal state value by inputting the internal state value and the external input value, and an output value generating means for converting the internal state value to the external output value. This is a characteristic speaker recognition method.

[0024]

EXAMPLES Example 1 FIG. 1 schematically shows the function of a nerve cell-like element that constitutes a neural network according to the present invention. In the figure, reference numeral 101 is an internal state value storage means of the nerve cell-like element, and 102 is an internal state value for updating the internal state value by inputting the internal state value stored in 101 and an external input value described below. The updating means, 103 is an output value generating means for converting an internal state value to an external output, and 104 is a schematic representation of the entire nerve cell-like element.

As the external input values shown in this figure, the output of the nerve cell-like element itself with a certain coupling weight added,
Similarly, the output of another neuron-like element with the coupling weights added thereto, the fixed output value with the coupling weights equivalently applied to bias the internal state updating means, and its neural network are input. Input data, etc. can be considered.

FIG. 2 schematically shows the function of a nerve cell-like element which constitutes a neural network by the conventional MLP method. In the figure, numeral 201 is an internal state value calculating means for calculating an internal state value, 202 is an output value generating means for converting the internal state value calculated by 201 into an external output, and 203 is an entire nerve cell-like element. Each is shown schematically.

As is clear from FIG. 2, the output value of the conventional nerve cell-like element is determined only by the input value at that time. In that sense, the operation of the conventional nerve cell-like element is static. In order for this static nerve cell-like device to process time series data, it is necessary to reflect the temporal structure of the target time series data in the structure of the neural network in some way or another.

On the other hand, in the neural network using the nerve cell-like element of the present invention, the past history of data is converted and held as the internal state value of the nerve cell-like element.
That is, the operation of the nerve cell-like device of the present invention is dynamic in the sense that the past history of the input is stored as this internal state value and is reflected in the output. Therefore, unlike the conventional neural network using a nerve cell-like element, the neural network of the present invention can process time series data regardless of the structure of the neural network.

As a modification of the conventional example, such history information may be stored as a context in a special neuron-like element group. However, in such a configuration, there is a problem that the function of the nerve cell-like element that constitutes the neural network becomes non-uniform and the processing becomes complicated. In any case, in the conventional technique, as described above as a problem, the process becomes complicated, the amount of data and the data memory increase, and the recognition accuracy decreases.

The operation of the nerve cell-like element that constitutes the present invention will be described in detail. The current internal state value is Xcur at each time change of the internal state value X and the output value Y.
r, the updated internal state value is Xnext, and the previously described external input value at the time of the updating operation is Zi (i is 0 to n, and n is the number of external inputs to the neuron-like element). , The operation of the internal state updating means is formally expressed as a function G, Xnext = G (Xcurr, Z1, ---, Zi,-
---, Zn) can be expressed. Although various concrete forms of this expression can be considered, for example, the following equation 2 using the first-order differential equation is used.
Something like is also possible. Where τ is a constant.

[0031]

[Equation 2]

Further, as a slightly modified form of this, an expression such as the following formula 3 is also possible.

[0033]

[Equation 3]

Among them, Wij represents the coupling strength for coupling the output of the j-th nerve cell-like element to the input of the i-th nerve cell-like element. Di represents an external input value. Also θ
i indicates a bias value. This bias value can be considered to be included in Wij as a combination with a fixed value.

When the internal state of the nerve cell-like element at a certain moment thus determined is X and the operation of the output value generating means is formally represented by a function F, the output Y of the nerve cell-like element is expressed.
Can be expressed as Y = F (X). As a concrete form of F, a sigmoid (logistic) function of positive / negative symmetrical output as shown by the following Expression 4 can be considered.

[0036]

[Equation 4]

However, this function type is not indispensable, and simpler linear conversion or a threshold function may be considered.

The time series of the output of the neural network according to the present invention is calculated by the processing as shown in FIG. In FIG. 7, the nerve cell-like element is simply described as a node for simplification.

Learning is required in order for the neural network to perform desired processing. Regarding this learning method, for example, the quantity C introduced by the following equation 5
There is a learning rule using.

[0040]

[Equation 5]

Here, C is a certain learning evaluation value and E is a certain error evaluation value. According to such a formula, C is shown in FIG.
It is determined by the processing as shown in.

As a concrete form of this error evaluation E, when the actual output value is Y and the desired output value is T, the following equation 6 is obtained.
The Kullback-leibler distance represented by is considered.

[0043]

[Equation 6]

When the output value range is between -1 and 1, it is substantially equivalent to the equation (6), but expressed as the following equation (7).

[0045]

[Equation 7]

Assuming these, the above equation 5 can be more specifically written as the following equation 8.

[0047]

[Equation 8]

By giving these, the updating rule of the weighting coefficient of various external inputs described above is given by the following Expression 9.

[0049]

[Equation 9]

In the present invention, unlike the back propagation learning in the conventional MLP method, the internal states and output values of all the nerve cell-like elements constituting the neural network can be updated at the same time. . Of course, it may be updated sequentially. Similarly, the configuration of the neural network in the present invention does not need to be layered as in the conventional MLP method. Fully-connected neural networks, layered neural networks, and other general-purpose neural networks are also possible.

FIG. 3 is a diagram schematically showing by what process the speaker feature time series data for speaker recognition is extracted. To explain the outline, first, the voice input by the voice input means is digitized by an AD converter or the like. After that, one part called a frame is extracted from the digitized input, and its features are extracted to form one feature vector. The temporal continuity of such a feature vector becomes the feature time series of the input speaker.

FIG. 5 is a schematic diagram of the configuration of a speaker recognition method according to an embodiment of the present invention. In the figure, reference numeral 501 is the speech feature extraction means as described in FIG. 3, 502 is the neural network as described above, and 504 is the input value storage means for storing the input of the neural network.
Reference numeral 03 schematically represents a data comparison means for comparing the stored input and output.

In this embodiment, the neural network is trained so that the feature vector corresponding to a certain frame is input and the feature vector of the immediately preceding frame is output. The speaker feature used here is the PA of the 8th order.
RCOR coefficient. PARCOR as the speaker feature
It is possible to use various other than the coefficient,
The PARCOR coefficient has such a characteristic that its value is in principle between -1 and 1, and has a relatively high rate of dependence on the speaker, which is a more effective characteristic amount in speaker recognition. is there.

In this embodiment, the input value storage means stores only the input data of one frame before, and compares it with the output data of the current frame to recognize the speaker. In other words, the speech feature time series, which can be easily understood from the mechanism of utterance, reflects the continuous movements of speech and articulatory organs. Since the individuality of the speaker appears as a feature of these movements, it is possible to recognize the individual by processing this feature time series. In this embodiment, the processing evaluates the individuality of the input voice by restoring the previous data from the current data. The structure of the neural network in this example is an asymmetric complete connection type structure including a self-loop. However, as described above, the neural network that constitutes the present invention can have a random configuration including layered connections and complete connections as special examples. The numbers of input elements, hidden elements, and output elements are all set.

In this embodiment, nine words are used as standard data for training the neural network,
"End point", "Prowess", "Rejection", "Transcendence", "For the time being", "Classification", "Rocker", "Mountains", and "Hidden Puritan" were used. Moreover, as the voice data, the one recorded in the Japanese voice database for research of the ATR person was used.

Further, according to the method of the present invention having the above-described configuration, it is possible to converge the learning found in the "BP model with feedback coupling" type neural network which is a modification of the conventional MLP method. The neural network of the speaker recognition method of the present invention is extremely easy to perform from several hundred times to several times, and there is no problem that the learning output for it is required to be created by trial and error. It was possible to generate the desired output after 1000 learnings.

9 and 10 show examples of results of speaker recognition by the neural network trained in this way. The solid line in the figure represents the time change of the error due to the output generated by the neural network trained to recognize the voice of the speaker MAU, and the broken line represents the neural network trained to recognize the voice of the speaker MXM. It shows the time change of the error of the generated output. The error shown here is a value obtained by averaging the absolute values of the lengths of the error vectors generated by the data comparison means with the 8th-order input vector data and the output vector for 32 frames before and after the frame at that time. It is shown. The input speaker in FIG. 9 is MAU, and the input speaker in FIG. 10 is MXM.

As is clear from the figure, in the case of FIG.
The data restoration error by the neural network trained by AU voice is small, and the restoration error by the neural network trained by MXM is larger. This is MA
It is shown that the data restoration using the utterance feature of U can perform the restoration more accurately, that is, the input voice is MAU.
It is due to.

Further, in the case of FIG. 10, contrary to the case of FIG.
The data restoration error due to the neural network trained by the XM voice is small, that is, the input voice is M
This is due to XM.

As is clear from the above figure, according to the speaker recognition method of the present invention, continuous speaker recognition results can be obtained.

Table 1 below shows the average value of the errors when a total of 11 voices including 9 speakers other than the training speaker are input to the two neural networks of the above example. The input was the 9 words that were used for training, and the average was performed over the entire utterance section. As is clear from the table, in each neural network, 11
The error with respect to the training speaker is the smallest with respect to the voice input of the person, and it is shown that the training speaker is correctly recognized from the 11 persons.

[0062]

[Table 1]

Table 2 below shows the same result as that of Table 1, but unlike the above case, it shows the result when a word voice having a different content from the word voice used for training is input. The words used here are "calendar", "welcome", "extreme", "parking", "program", "record", "purchase", and "typuter".

[0064]

[Table 2]

As is clear from the above table, it is shown that the speaker recognition method of the present invention accurately recognizes the training speaker even if the utterance content of the input voice is different.

Further, although the above description has explained the case of being discrete in time, it is also applicable to continuous time processing by performing analog processing, for example.

(Embodiment 2) FIG. 4 is a modification of Embodiment 1 in which training is carried out so that the input speech feature itself is output. In the figure, reference numeral 401 is a voice feature input means, 402 is a neural network, and 403 is a data comparison means.

Also in this example, the same effect as that of the first embodiment can be obtained.

(Third Embodiment) FIG. 6 shows a modification of the first embodiment in which training is performed so as to predict and output input data of a frame n future frames from the input speech feature. Reference numeral 601 in the figure denotes a voice feature input means 602.
Is a neural network, 603 is a data comparison means, and 604 is an output value storage means.

Also in this example, the same effect as that of the first embodiment can be obtained.

(Fourth Embodiment) FIG. 11 is an example different from the above embodiment in which the similarity between the input speaker and the speaker used for training is learned so as to be directly output. The input in this case can be the same as the one in the above embodiment, and as the learning output, for the input of a specific speaker, a specific similarity output element associated with that speaker is used. It should output the output. This output can be any number.

(Embodiment 5) FIG. 12 is similar to Embodiment 4 above. In this case, in addition to the fact that the specific similarity output element associated with a specific speaker outputs an output in response to the input of a specific speaker in the fourth embodiment, a speaker other than the target speaker is output. This is an example in which the input is learned so that the dissimilarity output outputs. This output can be any number as above. Generally, the dissimilarity output obtained as a result of such learning is not a simple inversion of the similarity output, and a higher degree of judgment can be made by combining them.

[0073]

As described above, according to the speaker recognition method of the present invention, 1) it is possible to perform highly accurate speaker recognition with a very small amount of learning data.

2) In order to recognize the utterance feature itself of the speaker, it is possible to recognize the speaker even if the data in the speaker recognition process is different from that in the training.

3) The learning is extremely easy, and there are very few trial and error parts for that. And so on.

The method of the present invention can be used not only for speaker recognition, but also for determining the degree of similarity between an unknown speaker and a known speaker. Further, the method of the present invention is effective not only for speech but also for processing of time series information in general.

[Brief description of drawings]

FIG. 1 is a schematic diagram of the function of a nerve cell-like element that constitutes a neural network in the present invention.

FIG. 2 is a schematic diagram of the function of a nerve cell-like element that constitutes a conventional neural network.

FIG. 3 is a schematic diagram of a configuration of voice feature extraction means.

FIG. 4 is a schematic diagram of an embodiment of the configuration of the speaker recognition method of the present invention.

FIG. 5 is a schematic diagram of one embodiment of the configuration of the speaker recognition method of the present invention.

FIG. 6 is a schematic diagram of an embodiment of the configuration of the speaker recognition method of the present invention.

FIG. 7 is a schematic diagram of a flow of processing of a neural network in the speaker recognition method of the present invention.

FIG. 8 is a schematic diagram showing a flow of error evaluation when learning a neural network in the speaker recognition method of the present invention.

FIG. 9 is a diagram showing a result of speaker recognition according to an embodiment of the present invention.

FIG. 10 is a diagram showing a result of speaker recognition according to an embodiment of the present invention.

FIG. 11 is a schematic diagram of an embodiment of the configuration of the speaker recognition method of the present invention.

FIG. 12 is a schematic diagram of one embodiment of the configuration of the speaker recognition method of the present invention.

[Explanation of symbols]

 101: internal state value storage means 102: internal state value updating means 103: output value generation means 104: nerve cell-like element 201: internal state value calculation means 202: output value generation means 203: nerve cell-like element 401: voice feature extraction Means 402: Neural network 403: Data comparison means 501: Voice feature extraction means 502: Neural network 503: Data comparison means 504: Input value storage means 601: Voice feature extraction means 602: Neural network 603: Data comparison means 604: Output value Storage unit 1101: Speech feature extraction unit 1102: Neural network 1103: Speaker similarity output 1 1104: Speaker similarity output 2 1201: Speech feature extraction unit 1202: Neural network 1203: Speaker similarity / dissimilarity output 1 1204 : Speaker similarity / dissimilarity output 2

Claims (12)

[Claims] 1. A speaker recognition method using a neural network, wherein the neural network updates at least an internal state value storage means and an internal state value by inputting an internal state value and an external input value. A speaker recognition method comprising a nerve cell-like element including: an output value generating means for converting an internal state value into an external output value. 2. The internal state value updating means of the nerve cell-like element, and the internal state value X of the nerve cell-like element, and the input Zi to the nerve cell-like element (i is 0 to n: n is a natural number). ), The following is obtained. The speaker recognition method according to claim 1, wherein the internal state value is updated to a value that satisfies 3. The speaker recognition method according to claim 1, wherein the speaker recognition method has a data comparison means for comparing data of an input value and an output value. . 4. The speaker recognition method has an input value storage means for storing an input value before a certain time T and a data comparison means for comparing the stored input value with the current output data. The speaker recognition method according to claim 1, wherein the speaker recognition method is a speaker recognition method. 5. The speaker recognition method has output value storage means for storing an output value before a certain time T, and data comparison means for comparing the stored output data with the current input data. The speaker recognition method according to any one of claims 1 and 2, which is characterized. 6. The speaker recognition method according to claim 1, wherein the speaker recognition method has a similarity output corresponding to each of one or more recognition target speakers. Speaker recognition method. 7. The speaker recognition method according to claim 1, wherein the speaker recognition method has a similarity output and a dissimilarity output corresponding to each of the one or more recognition target speakers. 2. The speaker recognition method according to any one of 2 above. 8. The input Zi is at least PARC.
8. The speaker recognition method according to claim 1, further comprising an OR coefficient. 9. The output value generating means is an output function having an output value between -1 and 1.
9. The speaker recognition method according to claim 8. 10. The speaker recognition method according to claim 1, wherein the input Zi includes at least a value obtained by multiplying an output of the nerve cell-like element itself by a weight. . 11. The speaker recognition according to claim 1, wherein the input Zi includes at least a value obtained by multiplying an output of another nerve cell-like element by a weight. Method. 12. The speaker recognition method according to claim 1, wherein the input Zi includes at least desired data provided from the outside.