KR20210015542A - Apparatus for identifying speaker based on in-depth neural network capable of enrolling unregistered speakers, method thereof and computer recordable medium storing program to perform the method - Google Patents

Abstract

A speaker identifying device based on a deep neural network capable of adding an unregistered speaker comprises: a deep neural network including an input layer and one or more hidden layers and output layers including a plurality of nodes, wherein a plurality of nodes of different layers are connected by weights; and a recognition unit identifying a speaker based on output values of a plurality of output nodes of the output layers after a speaker inputs an unknown voice to the input layer of the deep neural network and classifying the speaker as an unregistered speaker if all the output values of the plurality of output nodes are less than a preset threshold.

Description

Apparatus for identifying speaker based on in-depth neural network capable of enrolling unregistered speaker identification device based on a deep neural network capable of adding an unregistered speaker, a method therefor, and a program for performing the method speakers, method thereof and computer recordable medium storing program to perform the method}

The present invention relates to a speaker identification technology, and more particularly, a deep neural network-based speaker identification device capable of adding an unregistered speaker, a method therefor, and a computer-readable recording medium in which a program for performing the method is recorded. About.

A speaker identification system using a deep neural network is learned to identify predefined speakers. With the development of deep neural network technology, it has become possible to construct a speaker identification system with excellent performance. However, due to the characteristics of the classification system using a deep neural network, a large overhead occurs because a complete reconfiguration of the system is required in the process of adding an unregistered speaker.

Korean Patent Laid-Open Patent No. 2015-0104111 published on September 14, 2015 (Name: Speaker recording and identification using artificial neural network-based sub-speech unit distinction)

It is an object of the present invention to add an unregistered speaker capable of adding a new speaker by calculating the weight of the last layer instead of reconfiguring the entire system on the assumption that a new speaker is added to a speaker identification system using a deep neural network. A speaker identification device based on a deep neural network, a method therefor, and a computer-readable recording medium in which a program for performing the method is recorded can be provided.

A speaker identification device based on a deep neural network capable of adding an unregistered speaker according to a preferred embodiment of the present invention to achieve the above-described object includes an input layer, one or more hidden layers, and an output layer each including a plurality of nodes. Including, a deep neural network in which a plurality of nodes of different layers are connected by weights, and a speech of an unknown speaker is input to the input layer of the deep neural network, and then the output values of the plurality of output nodes of the output layer are based. And a recognition unit for classifying the speaker as an unregistered speaker when all of the output values of the plurality of output nodes are less than a preset threshold.

When classified as the non-registered speaker, an output node corresponding to the non-registered speaker is added to the output layer, and a plurality of hidden nodes of the last hidden layer and the output layer are added to the output layer based on the node value of the last hidden layer of the non-registered speaker stored at least a predetermined number of times. It further includes a learning unit that calculates the weight between the added output nodes.

The activation function of the output node is a softmax function, and the weight between the plurality of hidden nodes of the last hidden layer and the added output node is an average of node values of each of the plurality of hidden nodes of the last hidden layer. To do.

The threshold value is characterized in that the largest value among expected values of training data used when learning a registered speaker.

In a deep neural network-based speaker identification method in which an unregistered speaker can be added according to a preferred embodiment of the present invention for achieving the above object, the recognition unit includes an input layer including a plurality of nodes, one or more hidden layers, and The step of inputting a voice of an unknown speaker into a deep neural network including an output layer, wherein a plurality of nodes of a plurality of different layers are connected by weights, and the recognition unit includes a plurality of layers of the deep neural network according to the input. Deriving output values of the plurality of output nodes of the output layer through a plurality of operations on which the weight is applied by the node of, and the recognition unit identifies a speaker based on the output values of the plurality of output nodes of the output layer, Determining whether all of the output values of the plurality of output nodes are less than a preset threshold, and when the determination result, if all of the output values of the plurality of output nodes are less than a preset threshold, the recognition unit classifies the speaker as an unregistered speaker. Includes steps.

The speaker identification method includes the step of classifying the speaker as an unregistered speaker, then adding an output node corresponding to the unregistered speaker to an output layer by a learning unit, and a plurality of the last hidden layers of the unregistered speaker stored by the learning unit a predetermined number or more. And calculating weights between the plurality of hidden nodes of the last hidden layer and the output nodes added to the output layer based on the node values of the hidden nodes of.

The activation function of the output node is a softmax function, and in the calculating of the weight, the average of the node values of each of the plurality of hidden nodes of the last hidden layer is stored and the plurality of hidden nodes of the last hidden layer and the output It is characterized by calculating the weight between the output nodes added to the layer.

The threshold value is characterized in that the largest value among expected values of training data used when learning a registered speaker.

A computer-readable recording medium on which a program for performing a speaker identification method based on a deep neural network in which an unregistered speaker can be added according to a preferred embodiment of the present invention to achieve the above object is recorded has a plurality of recognition units. Inputting an unknown speaker to a deep neural network in which a plurality of nodes of a plurality of different layers are connected by weights, including an input layer including nodes of, at least one hidden layer, and an output layer, and the recognition unit Deriving output values of the plurality of output nodes of the output layer through a plurality of operations in which the weight is applied by a plurality of nodes of the plurality of layers of the deep neural network according to the input, and the recognition unit Identifying a speaker based on the output values of the output nodes, but determining whether all output values of the plurality of output nodes are less than a preset threshold, and as a result of the determination, all output values of the plurality of output nodes are less than a preset threshold. In this case, the recognition unit includes the step of classifying the speaker as an unregistered speaker.

The speaker identification method includes the step of classifying the speaker as an unregistered speaker, then adding an output node corresponding to the unregistered speaker to an output layer by a learning unit, and a plurality of the last hidden layers of the unregistered speaker stored by the learning unit a predetermined number or more. And calculating weights between the plurality of hidden nodes of the last hidden layer and the output nodes added to the output layer based on the node values of the hidden nodes of.

The activation function of the output node is a softmax function, and in the calculating of the weight, the average of the node values of each of the plurality of hidden nodes of the last hidden layer is stored and the plurality of hidden nodes of the last hidden layer and the output It is characterized by calculating the weight between the output nodes added to the layer.

The threshold value is characterized in that the largest value among expected values of training data used when learning a registered speaker.

According to the present invention, it is possible to add a speaker that can be recognized by calculating a weight between the last hidden layer and an output node corresponding to the added speaker, without relearning the entire deep-depth network. Accordingly, the computational complexity is reduced and the load on the system can be reduced.

1 is a block diagram illustrating an apparatus for identifying a speaker based on a deep neural network to which an unregistered speaker can be added according to an embodiment of the present invention.
2 is a view for explaining the configuration of a deep neural network according to an embodiment of the present invention.
3 is a diagram for explaining a node that performs an operation to which a weight is applied according to an embodiment of the present invention.
4 is a flowchart illustrating a method of learning a deep neural network according to an embodiment of the present invention.
5 is a flowchart illustrating a speaker recognition method using a deep neural network according to an embodiment of the present invention.
6 and 7 are flowcharts illustrating a method of adding an unregistered speaker according to an embodiment of the present invention.
8 is a diagram illustrating a method of adding an unregistered speaker according to an embodiment of the present invention.

Prior to the detailed description of the present invention, terms or words used in the present specification and claims described below should not be construed as being limited to their usual or dictionary meanings, and the inventors shall use their own invention in the best way. For explanation, based on the principle that it can be appropriately defined as a concept of terms, it should be interpreted as a meaning and concept consistent with the technical idea of the present invention. Therefore, the embodiments described in the present specification and the configurations shown in the drawings are only the most preferred embodiments of the present invention, and do not represent all the technical spirit of the present invention, and various equivalents that can replace them at the time of application It should be understood that there may be water and variations.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this case, it should be noted that the same components in the accompanying drawings are indicated by the same reference numerals as possible. In addition, detailed descriptions of known functions and configurations that may obscure the subject matter of the present invention will be omitted. For the same reason, some components in the accompanying drawings are exaggerated, omitted, or schematically illustrated, and the size of each component does not entirely reflect the actual size.

First, an apparatus for identifying a speaker based on a deep neural network to which an unregistered speaker can be added according to an embodiment of the present invention will be described. 1 is a block diagram illustrating an apparatus for identifying a speaker based on a deep neural network to which an unregistered speaker can be added according to an embodiment of the present invention. 2 is a view for explaining the configuration of a deep neural network according to an embodiment of the present invention. And FIG. 3 is a diagram for explaining a node that performs an operation to which a weight is applied according to an embodiment of the present invention.

First, referring to FIG. 1, an apparatus for identifying a speaker based on a deep neural network to which an unregistered speaker can be added according to an embodiment of the present invention (100, hereinafter, abbreviated as an identification device) is a communication unit 110, an audio system. It includes a unit 120, an input unit 130, a display unit 140, a storage unit 150, and a control unit 160.

The communication unit 110 is a means for communicating with other devices. The communication unit 110 may collect learning data from other devices through a network. The communication unit 110 may include a radio frequency (RF) transmitter Tx for up-converting and amplifying a frequency of a transmitted signal, and an RF receiver Rx for low-noise amplifying and down-converting a received signal. Further, the communication unit 110 may include a modem that modulates a transmitted signal and demodulates a received signal.

The audio unit 120 includes a speaker (SPK) for outputting an audio signal such as voice according to an embodiment of the present invention, and a microphone (MIKE) for collecting an audio signal such as voice. That is, the audio unit 120 may output an audio signal through the speaker SPK or transmit an audio signal input through the microphone (MIKE) to the controller 160 under the control of the controller 160.

The input unit 130 may receive a user's key manipulation for controlling the recognition device 10, generate an input signal, and transmit the input signal to the control unit 160. The input unit 130 may include various types of keys for controlling the recognition device 10. When the display unit 140 is formed of a touch screen, the input unit 130 may perform functions of various keys on the display unit 140, and when all functions can be performed only with the touch screen, the input unit 130 will be omitted. May be.

The display unit 140 is for displaying various types of information on a screen. In particular, the display unit 140 may output a guide for inputting an identifier to an unregistered speaker. In addition, the display unit 140 may visually provide a menu of the recognition device 10, input data, function setting information, and various other information to the user. In addition, the display unit 140 performs a function of outputting screens such as a boot screen, a standby screen, a menu screen, and the like of the recognition device 10. The display unit 140 may be formed of a liquid crystal display (LCD), an organic light emitting diode (OLED), an active matrix organic light emitting diode (AMOLED), or the like. Meanwhile, the display unit 140 may be implemented as a touch screen. In this case, the display unit 140 includes a touch sensor. The touch sensor detects a user's touch input. The touch sensor may be composed of a touch sensing sensor such as a capacitive overlay, a pressure type, a resistive overlay, or an infrared beam, or may be composed of a pressure sensor. . In addition to the above sensors, all types of sensor devices capable of sensing a contact or pressure of an object may be used as the touch sensor of the present invention. The touch sensor senses a user's touch input, generates a detection signal, and transmits it to the control unit 160. In particular, when the display unit 140 is formed of a touch screen, some or all of the functions of the input unit 130 are the display unit 140 It can be done through.

The storage unit 150 serves to store programs and data necessary for the operation of the recognition device 10. In particular, the storage unit 150 may store the node value of the last hidden node of the speaker, which is determined as an unregistered speaker, each time a voice of the corresponding speaker is input. Each type of data stored in the storage unit 160 may be deleted, changed, or added according to a user's manipulation.

The controller 160 may control the overall operation of the recognition device 10 and a signal flow between internal blocks of the recognition device 10 and perform a data processing function of processing data. In addition, the control unit 160 basically controls various functions of the recognition device 10. The control unit 160 includes a central processing unit (CPU) and a digital signal processor (DSP). Processor), etc. In particular, the control unit 160 includes a deep neural network 200 (DNN: Deep Neural Network), a learning unit 300 and a recognition unit 400. Such a deep neural network 200, The operation of the controller 160 including the learning unit 300 and the recognition unit 400 will be described in more detail below.

Then, the deep neural network 200 according to an embodiment of the present invention will be described in more detail with reference to FIG. 2. The deep neural network 200 according to an embodiment of the present invention includes a plurality of layers (IL, HL, OL). The plurality of layers includes an input layer (IL), at least one hidden layer (HL1 to HLk), and an output layer (OL).

In addition, each of the plurality of layers IL, HL, and OL includes a plurality of nodes. For example, as illustrated, the input layer IL may include n nodes i1 to in, and the output layer OL may include t nodes o1 to ot. In addition, among the hidden layers HL, the first hidden layer HL1 may include a nodes h11 to h1a, and the k-th hidden layer HLk may include c nodes hk1 to hkc.

All of a plurality of nodes of a plurality of hierarchies have operations. In particular, a plurality of nodes of different hierarchies are connected by channels (indicated by dotted lines) having weights (W). In other words, the weight is applied to the operation result of any one node and becomes the input of the next layer node. This connection relationship will be described with reference to FIG. 3. In FIG. 3, a node D, which is one node existing in the deep neural network 200, is illustrated as an example. Node D is the input signal x=[x1, x2, ... , xn] with weights w=[w1, w2,… , wn], then take the function F on the result. Here, the function F is referred to as an activation function or a transfer function. In this case, even when the inputs are the same, the outputs have different values according to the weight W.

That is, the output of each node is shown in Equation 1 below.

의 값이 임계치 보다 작을 때 해당 노드가 활성화되지 않도록 하는 역할을 한다. Unexplained variable θ is a threshold or bias, and this threshold is

When the value of is less than the threshold, the node is not activated.

For example, assume that there are three nodes in the previous layer of any one node D. Accordingly, there are three inputs (n=3) X1, X2, X3 and three weights W1, W2, and W3 for the node.

Node D receives a value obtained by multiplying the weights W1, W2, and W3 corresponding to the three inputs X1, X2, X3, sums them all, and calculates the output by substituting the summed value into the transfer function. Specifically, it is assumed that the input [X1, X2, X3] = 0.5, -0.3, 0, and the weight w = [W1, W2, W3] = 4, 5, 2. Also, for convenience of explanation, assuming that the transfer function is'sgn()', the output value is calculated as follows.

x1 × w1 = 0.5 × 4 = 2

x2 × w2 =-0.3 × 5 = -1.5

x3 × w3 = 0 × 2 = 0

2 + (-1.5) + 0 = 0.5

sgn(0.5) = 1

In this way, any one node of any one layer of the deep neural network 200 receives a value obtained by applying a weight to the input from the node of the previous layer, summing it to take a transfer function, and converts the result to the input of the next layer. Deliver.

However, a softmax function is used as the activation function or transfer function of a plurality of output nodes of the output layer. The softmax function normalizes all input values to output values between 0 and 1, and has a characteristic that the sum of output values is always 1. This soft pulse function is shown in Equation 2 below.

Here, t represents the output node of the output layer.

On the other hand, returning to FIG. 2, the user's voice is input to the input layer IL, and according to this input, a plurality of nodes of the hidden layer HL perform an operation as described in FIG. 3 and calculate the result of the operation. Pass it to the output layer (OL). Then, each node of the output layer OL outputs an output value.

In FIG. 2, t nodes o1 to ot of the output layer OL correspond to t different speakers. That is, each of the first to tth output nodes corresponds to the first to tth speakers. When the voice of one of the t speakers is input to the input layer IL of the deep neural network 200, if the output value of the second output node o2 is greater than the output value of the other output nodes o1, o3 ~ ot , The corresponding voice represents the second speaker. That is, when a voice is input to the deep neural network 200, if the output value of one of the plurality of output nodes is greater than the output value of the other output nodes, the corresponding voice is the speaker corresponding to the output node having the largest output value. Judging by voice.

In order for the deep neural network 200 to recognize as described above, machine learning is required. Learning according to an embodiment of the present invention is a procedure in which a speaker uses a known speech as learning data, sets an expected value, and adjusts the weight w of a node of the deep neural network 200 according to the expected value.

This learning will be described in more detail with reference to FIG. 4. 4 is a flowchart illustrating a method of learning a deep neural network according to an embodiment of the present invention. 2 to 4, in this embodiment, it is assumed that there are two output nodes (t=2), and there are a first speaker and a second speaker. In addition, it is assumed that each of the first speaker and the second speaker corresponds to the first output node o1 and the second output node o2. The voice of the first speaker is used for learning to identify the first speaker, and the voice of the second speaker is used for learning to identify the second speaker. In other words, the learning data uses a voice known by the speaker.

First, the learning unit 300 sets an expected value according to the speaker to be trained in step S110. For example, the learning data for the first speaker, that is, the expected value for the voice of the first speaker, is, for example, the first output. The output value of the node o1 may be 0.8 or more, and the output value of the second output node o2 may be set to 0.2 or less (o1≥0.8, o2≤0.2) In addition, learning data about the second speaker, that is, the The expected value for the voice of the second speaker may be set as, for example, the output value of the first output node o1 is 0.2 or less, and the output value of the second output node o2 is 0.8 or more (o1≤0.2, o2≥0.8). have.

Next, the learning unit 300 inputs the corresponding learning data into the deep neural network 200. The deep neural network 200 performs an operation as described in FIG. 3 in a plurality of layers and a plurality of nodes, and the result The phosphorus output value will be output through the output nodes o1 and o2, which may differ from the previously set expected value.

Then, the learning unit 300 calculates the difference between the expected value and the output value in step S130, and adjusts the weight W of each node through a back propagation algorithm so that the difference is minimized in step S140. In other words, The learning unit 300 adjusts the weight (W) of each node so that the output value becomes the expected value. Adjusting the weight (W) of each node is called learning, and such learning is insufficient in the above-described one-time process. It is preferable to perform iteratively until an output value always satisfies an expected value using a plurality of training data.

In summary, learning is a process of comparing an output value and a preset expected value when training data is actually input to the deep neural network 200, and adjusting the weight w so that the output value becomes the expected value. More specifically, referring to FIG. 3 again, it is assumed that the node D is one of the output nodes. When input [x1, x2, x3] = 0.5, -0.3, 0, weight w=[w1, w2, w3] = 4, 5, 2, and transfer function is'sgn()', output of node N Was 1. At this time, it is assumed that the expected value of node D is 0. Then, w2 can be modified from 5 to 6 by learning, and the output value becomes the expected value as follows.

x1 × w1 = 0.5 × 4 = 2

x2 × w2 =-0.3 × 6 = -1.8

x3 × w3 = 0 × 2 = 0

2 + (-1.8) + 0 = 0.2

sgn(0.2) = 0

When the above-described learning is repeatedly performed using a plurality of different learning data having the same expected value, it is preferable to repeatedly perform the learning until the output value satisfies the expected value without changing the output value. In this way, learning, a procedure of adjusting the weights of the deep neural network so that the difference between the expected value and the output value is minimal, includes a procedure of modifying the weights for a plurality of nodes in reverse order from the output layer to the input layer through a despreading algorithm. do.

When learning is completed, the deep neural network 200 may recognize the speaker. Then, a speaker recognition method according to an embodiment of the present invention will be described. 5 is a flowchart illustrating a speaker recognition method using a deep neural network according to an embodiment of the present invention. 2 to 5, in this embodiment, it is assumed that there are two output nodes (t=2), and the first speaker and the second speaker are in a learned state. In addition, it is assumed that each of the first speaker and the second speaker corresponds to the first output node o1 and the second output node o2.

Referring to FIG. 5, the recognition unit 400 inputs a voice of an unknown speaker into the deep neural network 200 in step S210. Then, the deep neural network 200 will output the output values through the plurality of output nodes o1 and o2 through a plurality of operations to which the weights of the plurality of layers are applied.

Therefore, the recognition unit 400 recognizes the speaker according to the output value in step S220. For example, it is assumed that the output values of each of the first and second output nodes are 0.85 and 0.15 (o1 = 0.85, o2 = 0.15). This indicates that the probability that the input voice is the voice of the first speaker is 85% and that the voice of the second speaker is 15%. Accordingly, the recognition unit 400 may recognize that the input voice is the voice of the first speaker.

Meanwhile, in the case of a learned speaker, it can be recognized that the corresponding speaker is a voice. However, if the voice of the speaker is not learned in advance, the deep neural network 200 cannot recognize the speaker. In this case, the present invention can add a new speaker. This method will be described. 6 and 7 are flowcharts illustrating a method of adding an unregistered speaker according to an embodiment of the present invention. 8 is a diagram for explaining a method of adding an unregistered speaker according to an embodiment of the present invention.

First, referring to FIG. 6, in this embodiment, it is assumed that there are two output nodes (t=2), and the first speaker and the second speaker are in a learned state. In addition, it is assumed that each of the first speaker and the second speaker corresponds to the first output node o1 and the second output node o2.

The recognition unit 400 inputs a voice of an unknown speaker into the deep neural network 200 in step S310, and derives an output value through the deep neural network 200 in step S320. Subsequently, the recognition unit 400 determines whether all of the output values of the output node are less than a preset threshold in step S330. In an embodiment of the present invention, the threshold may be the largest value among expected values of training data used when learning a registered speaker. As described above, the expected value of the training data for the first speaker is, for example, the output value of the first output node o1 is 0.8 or more, and the output value of the second output node o2 is 0.2 or less (o1≥0.8, o2≤ 0.2), and the expected value of the training data for the second speaker is, for example, the output value of the first output node o1 is 0.2 or less, and the output value of the second output node o2 is 0.8 or more (o1≤0.2, o2). ≥0.8). In this case, the threshold may be 0.8.

If all of the output values of the output nodes o1 and o2 are less than a preset threshold (e.g., 0.8) (e.g., o1 = 0.7, o2 = 0.3), the recognition unit 400 selects the corresponding speaker as an unregistered speaker in step S340. Classify (determining that the corresponding voice is the voice of an unregistered speaker), and perform a speaker addition process. On the other hand, if all of the output values of the output nodes are not less than the preset threshold, the recognition unit 400 recognizes the speaker according to the output value of the deep neural network 200 in step S350. This step S350 is the same as step S220 described above.

Then, a process of adding a speaker will be described with reference to FIG. 7. If all of the output values of the output nodes o1 and o2 are less than a preset threshold (eg, o1 = 0.7, o2 = 0.3), the learning unit 300 receives the identifier of the corresponding unregistered speaker in step S410. For example, the learning unit 300 may guide the input of the speaker's name (identifier) through the audio unit 120 and the display unit 140 to receive an identifier of an unregistered speaker.

When the identifier of the unregistered speaker is input, the learning unit 300 maps the identifier of the unregistered speaker input in step S420 to store the node values of the plurality of hidden nodes of the last hidden layer. Referring to FIG. 8, it is assumed that k=2, for example. Then, the last hidden layer may be the second hidden layer. Then, the learning unit 300 maps to the inputted identifier of the unregistered speaker and stores the node values of each of the plurality of hidden nodes (hk1 to hkc, k=2) of the last hidden layer (the second hidden layer). Save it to.

Subsequently, in step S430, the learning unit 300 determines whether the number of storing the node value of the last hidden layer corresponding to the identifier of the corresponding unregistered speaker is equal to or greater than a preset number. This step S430 is to determine whether or not the number of learning data for sufficiently learning the voice of the corresponding non-registered speaker has been secured.

If, as a result of the determination in step S430, the node value of the last hidden layer corresponding to the corresponding unregistered speaker is stored more than a preset number of times, the learning unit 300 adds an output node in step S440. For example, referring to FIG. 8 compared to FIG. 2, the learning unit 300 adds the t+1th output node ot+1 to the output layer. The added t+1th output node ot+1 corresponds to a newly registered unregistered speaker.

Next, the learning unit 300 uses the node values of the plurality of hidden nodes of the last hidden layer corresponding to the corresponding unregistered speaker stored in the storage unit 150 in step S450, and the plurality of hidden nodes and output layers of the last hidden layer. The weight (W_add) between output nodes added to is calculated.

Since the active function of the output layer is a soft max function, the weight matrix connecting the last hidden layer and the output layer is a two-dimensional matrix of the number of dimensions of the last hidden layer x the number of registered speakers. In order to add a new unregistered speaker as a new speaker to the weight matrix, a vector of the number of dimensions of the last hidden layer representing the t+1th speaker is added. Specifically, the vector representing the t+1th speaker uses the average of the node values of the hidden nodes of the last hidden layer when the previously stored voice of the t+1th speaker is input. As a result, the weight matrix has a size of (the number of dimensions of the last hidden layer x the number of registered speakers + 1), and when a new utterance is input thereafter, speaker identification can be performed including the t+1th speaker.

Meanwhile, as a result of the determination in step S330, if the node value of the last hidden layer corresponding to the corresponding non-registered speaker is stored less than a preset number of times, the corresponding process is terminated because the corresponding non-registered speaker cannot be registered.

In addition, the method according to the embodiment of the present invention described above may be implemented in the form of a program that can be read through various computer means and recorded in a computer-readable recording medium. Here, the recording medium may include a program command, a data file, a data structure, or the like alone or in combination. The program instructions recorded on the recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. For example, the recording medium includes magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic-optical media such as floptical disks ( magneto-optical media), and hardware devices specially configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of the program instruction may include not only machine language wires such as those made by a compiler, but also high-level language wires that can be executed by a computer using an interpreter or the like. Such a hardware device may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

The present invention has been described above using several preferred embodiments, but these embodiments are illustrative and not limiting. As such, those of ordinary skill in the art to which the present invention pertains will understand that various changes and modifications can be made according to the equivalence theory without departing from the spirit of the present invention and the scope of the rights presented in the appended claims.

100: identification device 110: communication unit
120: audio unit 130: input unit
140: display unit 150: storage unit
160: control unit 200: neural network
300: learning unit 400: recognition unit

Claims (12)

In a deep neural network-based speaker identification device capable of adding an unregistered speaker,
A deep neural network, each including an input layer including a plurality of nodes, at least one hidden layer, and an output layer, wherein a plurality of nodes of different layers are connected by weights; And
After the speaker inputs an unknown voice to the input layer of the deep neural network, the speaker is identified based on the output values of a plurality of output nodes of the output layer, but if all output values of the plurality of output nodes are less than a preset threshold, And a recognition unit for classifying the speaker as an unregistered speaker.
Speaker identification device. The method of claim 1,
When classified as the non-registered speaker, an output node corresponding to the non-registered speaker is added to the output layer, and a plurality of hidden nodes of the last hidden layer and the output layer are added to the output layer based on the node value of the last hidden layer of the non-registered speaker stored at least a predetermined number of times. Characterized in that it further comprises a; learning unit for calculating the weight between the added output nodes
Speaker identification device. The method of claim 2,
The activation function of the output node is a softmax function,
The weight between the plurality of hidden nodes of the last hidden layer and the added output node is an average of node values of each of the plurality of hidden nodes of the last hidden layer.
Speaker identification device. The method of claim 1,
The threshold value is the largest value among expected values of training data used when learning a registered speaker.
Speaker identification device. In a deep neural network-based speaker identification method that can add an unregistered speaker,
A step of inputting an unknown speaker to a deep neural network in which the recognition unit includes an input layer each including a plurality of nodes, at least one hidden layer, and an output layer, and a plurality of nodes of different layers are connected by weights ;
Deriving, by the recognition unit, output values of a plurality of output nodes of an output layer through a plurality of operations in which the weight is applied by a plurality of nodes of a plurality of layers of the deep neural network according to the input;
Identifying a speaker based on output values of the plurality of output nodes of the output layer by the recognition unit, and determining whether all output values of the plurality of output nodes are less than a preset threshold; And
And classifying, by the recognition unit, the speaker as an unregistered speaker if all of the output values of the plurality of output nodes are less than a preset threshold as a result of the determination.
How to identify a speaker. The method of claim 5,
After the step of classifying the speaker as an unregistered speaker,
Adding, by a learning unit, an output node corresponding to the unregistered speaker to an output layer;
Calculating weights between the plurality of hidden nodes of the last hidden layer and the output nodes added to the output layer based on the node values of the plurality of hidden nodes of the last hidden layer of the unregistered speaker stored at least a predetermined number of times by the learning unit; Characterized in that it comprises
How to identify a speaker. The method of claim 6,
The activation function of the output node is a softmax function,
The step of calculating the weight is
And calculating the average of the node values of each of the plurality of hidden nodes of the last hidden layer stored as a weight between the plurality of hidden nodes of the last hidden layer and the output node added to the output layer.
How to identify a speaker. The method of claim 5,
The threshold value is the largest value among expected values of training data used when learning a registered speaker.
How to identify a speaker. In a computer-readable recording medium in which a program for performing a speaker identification method based on a deep neural network capable of adding an unregistered speaker is recorded,
Step of inputting an unknown speaker to a deep neural network in which the recognition unit includes an input layer, each of which includes a plurality of nodes, at least one hidden layer, and an output layer, and a plurality of nodes of a plurality of different layers are connected by weights ;
Deriving, by the recognition unit, output values of a plurality of output nodes of an output layer through a plurality of operations in which the weight is applied by a plurality of nodes of a plurality of layers of the deep neural network according to the input;
Identifying a speaker based on output values of the plurality of output nodes of the output layer by the recognition unit, and determining whether all output values of the plurality of output nodes are less than a preset threshold; And
As a result of the determination, if all of the output values of the plurality of output nodes are less than a preset threshold, the recognition unit classifying the speaker as an unregistered speaker;
A computer-readable recording medium on which a program for performing a speaker identification method is recorded. The method of claim 9,
After the step of classifying the speaker as an unregistered speaker,
Adding, by a learning unit, an output node corresponding to the unregistered speaker to an output layer;
Calculating weights between the plurality of hidden nodes of the last hidden layer and the output nodes added to the output layer based on the node values of the plurality of hidden nodes of the last hidden layer of the unregistered speaker stored at least a predetermined number of times by the learning unit; Included
A computer-readable recording medium on which a program for performing a speaker identification method is recorded. The method of claim 10,
The activation function of the output node is a softmax function,
The step of calculating the weight is
And calculating the average of the node values of each of the plurality of hidden nodes of the last hidden layer stored as a weight between the plurality of hidden nodes of the last hidden layer and the output node added to the output layer.
A computer-readable recording medium on which a program for performing a speaker identification method is recorded. The method of claim 9,
The threshold value is the largest value among expected values of training data used when learning a registered speaker.
A computer-readable recording medium on which a program for performing a speaker identification method is recorded.

Images