KR20200107707A - Registration method and apparatus for speaker recognition - Google Patents

Abstract

Disclosed are a registration method and device for speaker recognition. The registration method for speaker recognition according to one embodiment of the present invention comprises the steps of: receiving a voice signal of a speaker; synthesizing a noise signal to the voice signal; generating a feature vector based on the synthesized signal of the voice signal and the noise signal; and building a registration database corresponding to the speaker based on the feature vector.

Description

Registration method and device for speaker recognition {REGISTRATION METHOD AND APPARATUS FOR SPEAKER RECOGNITION}

The following embodiments relate to a registration method and apparatus for speaker recognition.

Speaker recognition technology is a technology that verifies the identity through the speaker's voice, and can be applied to various fields such as identification in meetings, conferences, and conversations. The speaker recognition system can also be used for controlling access to automobiles, buildings, bank accounts, etc. for security purposes, and the need for development thereof has emerged. In order to authenticate a speaker with the speaker's voice, a strong response to noise is required, and a method is needed to make the most of the limited speaker's voice information.

A registration method for speaker recognition according to an embodiment includes the steps of receiving a speech signal of a speaker; Synthesizing a noise signal to the speech signal; Generating a feature vector based on a synthesized signal of the speech signal and the noise signal; And constructing a registration database corresponding to the speaker based on the feature vector.

The registration method according to an embodiment further comprises generating a second feature vector based on the voice signal, and the step of constructing the registration database is based on at least one of the feature vector and the second feature vector. It may include the step of building the registration database.

Generating the feature vector may include domain transforming the synthesized signal; And extracting the feature vector from the domain-transformed result.

The synthesizing may include adding an additive noise signal to the speech signal.

The synthesizing may include performing a convolution operation of the speech signal and a channel noise signal.

The additive noise signal may be determined based on at least one of a noise type, a timing of noise, and an energy ratio (SNR) of noise to the voice signal.

The channel noise signal may be determined based on at least one of a noise type and an energy ratio (SNR) of noise to the voice signal.

The generating of the feature vector may include segmenting a synthesized signal of the speech signal and the noise signal into a plurality of sections; And extracting at least one section feature vector corresponding to at least one of the plurality of sections.

The registration method according to an embodiment includes the steps of constructing the registration database comprising: clustering a plurality of feature vectors including at least one of the feature vector and the second feature vector into a plurality of groups; The method may further include extracting at least one representative feature vector corresponding to at least one of the plurality of groups.

The registration method according to an embodiment further comprises extracting a starting point and an ending point of sound information included in at least one of the speech signal and the synthesized signal, and synthesizing the starting point and the ending point. And at least one of the generating steps.

The synthesized signal includes sound information in a time domain, the domain transformed result includes image information in a frequency domain, and generating the feature vector is performed using a convolutional neural network (CNN). It may include the step of extracting the feature vector from the image information in the frequency domain.

When receiving the speaker's voice signal through a plurality of channels in the receiving step, the synthesizing step and the generating step are independently performed corresponding to each of the plurality of channels, and the plurality of The registration database may be constructed based on feature vectors generated by two channels.

A registration method for speaker recognition according to an embodiment includes, after the receiving step, estimating noise based on differences between the speaker's voice signals received through the plurality of channels; And subtracting the noise from each of the speaker's voice signals.

The receiving of the speaker's voice signal includes receiving the speaker's voice signals through a plurality of channels, and the registration method for speaker recognition according to an embodiment includes the plurality of channels. The method may further include determining the voice signal for the synthesizing step by preprocessing the received voice signals of the speaker.

The determining may include selecting any one of the speech signals based on whether a recognition error due to a hardware spec is included in the speech signals of the speaker received through the plurality of channels; It may include the step of selecting any one of the voice signals for each section based on whether the recognition error is included for each section of the speaker's voice signals received through the plurality of channels.

A speaker recognition method according to an embodiment includes the steps of: receiving a speaker's voice signal; Generating at least one input feature vector based on the speech signal; Comparing at least one registered feature vector of a registered user stored in a registration database with the at least one input feature vector; And recognizing the speaker based on the comparison result, wherein the registration database includes the at least one registration feature vector generated based on a synthesized signal of a voice signal and a noise signal for registration of the registered user. Include.

When the number of registered users is plural, recognizing the speaker may include identifying the speaker.

The generating of the input feature vector may include domain conversion of the speaker's speech signal; And extracting the input feature vector from the domain-transformed result.

The generating of the input feature vector may include performing a sliding window on the speaker's speech signal so that the window size of the input feature vector is the same as the window size of the registered feature vector.

The method of claim 12, wherein the number of input feature vectors may be determined based on a security level or response time required by an application.

Domain transforming may include performing a fast Fourier transform on the speaker's voice signal; And performing padding on the result of the fast Fourier transform according to the result of the fast Fourier transform.

The predetermined noise signal may include at least one of an additive noise signal and a channel noise signal.

A registration device for speaker recognition according to an embodiment includes: an acoustic sensor for receiving a speech signal of a speaker; A processor for synthesizing a noise signal with the speech signal, generating a feature vector based on the synthesized signal of the speech signal and the noise signal, and constructing a registration database corresponding to the speaker based on the feature vector. Include.

The processor may generate a second feature vector based on the speech signal, and build the registration database based on at least one of the feature vector and the second feature vector.

The processor may perform domain transformation on the synthesized signal and extract the feature vector from the domain transformation result.

The processor may add an additive noise signal to the voice signal.

The processor may perform a convolution operation of the voice signal and a channel noise signal.

The processor may segment the synthesized signal of the speech signal and the noise signal into a plurality of sections, and extract at least one section feature vector corresponding to at least one of the plurality of sections.

The processor clusters a plurality of feature vectors including at least one of the feature vector and the second feature vector into a plurality of groups, and at least one representative feature vector corresponding to at least one of the plurality of groups Can be extracted.

A speaker recognition apparatus according to an embodiment includes an acoustic sensor for receiving a speaker's voice signal; And generating at least one input feature vector based on the voice signal, comparing at least one registered feature vector of a registered user stored in a registration database with the at least one input feature vector, and the speaker based on the comparison result. And a processor for recognizing, wherein the registration database includes the at least one registration feature vector generated based on a synthesized signal of a voice signal and a noise signal for registration of the registered user.

The processor may perform domain conversion of the speaker's speech signal and extract the input feature vector from the domain conversion result.

The processor may perform a sliding window on the speaker's speech signal so that the window size of the input feature vector is the same as the window size of the registered feature vector.

The processor may perform fast Fourier transform on the speaker's voice signal, and may perform padding on the fast Fourier transform result according to the fast Fourier transform result.

1 is a diagram for describing a speaker recognition method according to an exemplary embodiment.
2 is a diagram illustrating a registration system for speaker recognition according to an exemplary embodiment.
3 is a flowchart illustrating a registration method for speaker recognition according to an embodiment.
4 is a diagram illustrating a method of generating a feature vector by synthesizing additive noise to a speech signal according to an exemplary embodiment.
5 is a diagram for describing a method of generating a feature vector by synthesizing channel noise with a voice signal according to an exemplary embodiment.
6 is a diagram illustrating a method of generating a feature vector by segmenting according to an exemplary embodiment.
7 is a diagram for describing a method of constructing a registration database according to an embodiment.
8 is a diagram for describing a speaker authentication method according to an embodiment.
9 is a flowchart illustrating a speaker recognition method according to an exemplary embodiment.
10 is a diagram for describing a speaker identification method according to an exemplary embodiment.
Fig. 11 is a block diagram of a registration device for speaker recognition according to an embodiment.
12 is a diagram illustrating a method of authenticating a speaker through a plurality of channels according to an exemplary embodiment.
13 is a diagram for explaining a speaker authentication method through channel selection according to an embodiment.
14 is a diagram for describing a method of overcoming a change in vocalization according to a location and direction of a registration device according to an exemplary embodiment.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are exemplified only for the purpose of describing the embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention They may be implemented in various forms and are not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention can apply various changes and have various forms, the embodiments will be illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes changes, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first or second may be used to describe various elements, but the elements should not be limited by the terms. The above terms are only for the purpose of distinguishing one component from other components, for example, without departing from the scope of rights according to the concept of the present invention, the first component may be named as the second component, Similarly, the second component may also be referred to as a first component.

When a component is referred to as being “connected” or “connected” to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in the middle. Should be. On the other hand, when a component is referred to as being “directly connected” or “directly connected” to another component, it should be understood that there is no other component in the middle. Expressions that describe the relationship between components, for example, “between” and “just between” or “directly adjacent to” should be interpreted as well.

The terms used in the present specification are only used to describe specific embodiments and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present specification, terms such as "include" or "have" are intended to designate as being implemented features, numbers, steps, actions, components, parts, or a combination thereof, but one or more other features or numbers, It is to be understood that the presence or addition of steps, actions, components, parts, or combinations thereof, does not preclude the possibility of preliminary exclusion.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this specification. Does not.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. The same reference numerals in each drawing indicate the same members.

1 is a diagram for describing a speaker recognition method according to an exemplary embodiment. Referring to FIG. 1, a method for recognizing a speaker according to an embodiment may be divided into a registration step and a test step. The speaker recognition method may be performed by a speaker recognition device, and the speaker recognition device may be implemented by one or more software modules, one or more hardware modules, or various combinations thereof.

According to the speaker recognition method according to an embodiment, in the registration step, the speaker recognition device may generate a feature vector from the speaker's input voice and store it in the registration database, and in the test step, the feature vector is generated from the speaker's input voice. It is possible to determine whether to recognize or not by comparing it with the feature vector created and stored in the registration database. For convenience of description, embodiments in which the registration step and the test step are performed by the same speaker recognition device are described. However, the embodiments described below are substantially performed even when the registration step and the test step are performed by different devices. The same can be applied.

The voice signal 110 may be a voice signal of a speaker corresponding to a predetermined recognition sentence to be authenticated. For example, the voice signal 110 may be a speaker's voice signal corresponding to the recognition sentence “hi, bixby”.

In the registration step, the speaker recognition apparatus may generate a feature vector by receiving a voice signal corresponding to the speaker's voice “hi, bixby”, and store the generated feature vector in the registration database. Thereafter, when receiving a voice signal corresponding to the speaker's voice "hi, bixby" in the test step, the speaker recognition apparatus may generate a feature vector and compare it with the feature vector stored in the registration database to determine whether to recognize it.

According to an embodiment, the registration step may be performed in a quiet environment with less noise, while the test step may be performed in a noisy environment. In this case, in the registration step, a feature vector may be generated based on the voice signal 120 in a low noise environment. On the other hand, in the test step, noise may be added to the speaker's voice signal according to the surrounding environment, and feature vectors may be generated based on the noise-added voice signals 130 and 140. The noise may include additive noise, channel noise, or a combination thereof. For example, a voice signal to which noise is added may include a voice signal 130 to which additive noise is added and a voice signal 140 to which channel noise is added. In this case, a discrepancy between the feature vector generated in the registration step and the feature vector generated in the test step may occur, resulting in a decrease in recognition performance. In addition, when the number of utterances of the speaker's speech signal corresponding to the recognition sentence is small in the registration step, an error is likely to occur in the feature vector, which may cause difficulty in recognition.

According to an embodiment, in order to prevent performance degradation caused by a mismatch between a registration environment and a test environment, a feature vector may be registered based on a synthesized signal obtained by synthesizing a predetermined noise signal with a speech signal of a speaker in the registration step. Through this, excellent performance can be secured during authentication in a real environment situation exposed to various noises different from the registered environment.

Looking at the classification of speaker recognition, speaker recognition technology can be divided into speaker verification and speaker identification. Speaker authentication is to authenticate whether the speaker is a previously registered user (eg, the person), and speaker identification may mean finding out who the speaker is among several registered users.

Speaker recognition technology can be widely applied from the management of personal electronic documents to banking and electronic commerce over the phone or the Internet. The reason why speaker recognition technology using voice is preferred is that there is no problem of theft or forgery of physical means such as cards, stamps, signatures, and ID cards, and it is expensive to input fingerprints or retinas, which are other biometric means. This is because voice can be easily processed from a distance through an acoustic sensor such as a microphone, which is relatively inexpensive, while the equipment is required.

In the following, a speaker recognition registration method strong against noise is described with reference to FIGS. 2 to 7, a speaker authentication method is described with reference to FIGS. 8 to 9, a speaker recognition method is described with reference to FIG. 10, and further A registration device and a recognition device according to an embodiment will be described with reference to FIG. 11.

2 is a diagram illustrating a registration system for speaker recognition according to an exemplary embodiment. Referring to FIG. 2, a registration device for speaker recognition may receive a speaker's voice signal 210. The registration device is a device that builds a registration database corresponding to a speaker for speaker recognition, and may be implemented as, for example, a software module, a hardware module, or a combination thereof. The registration device may be mounted on various computing devices and/or systems, such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a television, a wearable device, a security system, and a smart home system.

The speaker's voice signal 210 may be a signal that has been input through an acoustic sensor of the registration device and has not been specially processed, or may be a signal that has been input through an acoustic sensor and a component corresponding to noise has been removed. Even in the case of a signal that has not been specially processed, since the registration step is generally performed in a quiet and low-noise environment, there may be substantially no difference from the signal from which the component corresponding to the noise has been removed.

In order to prevent performance degradation caused by inconsistency between the registration environment and the test environment, the registration device may synthesize the additive noise signal 220, the channel noise signal 230, or a combination thereof with the voice signal. Additive noise may be noise that arises from a different source than the speech signal and has no correlation between them. Additive noise can be added to the speech signal by addition. For example, additive noise may include a door closing sound, a horn sound, and ambient noise. Channel noise may be the type of noise found in the conversion process. For example, the channel noise may include a room impulse response. However, additive noise and channel noise are exemplary only, and noise may include various types of noise in addition to additive noise and channel noise. A detailed method of synthesizing the additive noise signal 220 and the channel noise signal 230 will be described in detail below with reference to FIGS. 4 and 5, respectively.

According to an embodiment, the registration device may arrange 240 voice signals. The segment to be processed may be determined by aligning the voice signal, and through this, the start point and the end point of the sound information included in the voice signal may be extracted to increase registration and recognition efficiency. For example, the registration device may measure the strength of the voice signal to extract the start point and the end point of sound information included in the voice signal, and set a section between the start point and end point of the sound information as a processing target section. I can. Here, the signal to be aligned may be a signal obtained by synthesizing the additive noise 220, the channel noise 230, or a combination thereof into a voice signal.

Although not shown in the drawings, according to another embodiment, the registration device first aligns the voice signals 210 to extract the start and end points of the sound information, and then the additive noise 220, the channel noise 230, or these A combination of may be synthesized into the voice signal 210. Alternatively, the voice signal 210 may be first aligned from other devices such as a digital signal processor and input to the registration device, and thereafter, additive noise 220, channel noise 230, or a combination thereof May be synthesized into the voice signal 210.

According to an embodiment, the registration device may segment a synthesized signal of a voice signal and a noise signal into a plurality of sections (segmentation) 250. A specific method for segmentation will be described in detail below with reference to FIG. 6.

The feature vector generator 260 of the registration device may generate a feature vector based on a synthesized signal of a voice signal and a noise signal. The feature vector generator 260 may recognize the synthesized signal and output a feature vector corresponding thereto. The feature vector includes information for distinguishing each recognition element. For example, the feature vector may have frequency information over time in a form in which components necessary for recognition are compressed and contained in a speech signal.

According to an embodiment, by generating and registering a feature vector based on a synthesized signal obtained by synthesizing various noise signals with a speaker's voice signal in the registration step, performance degradation caused by a discrepancy between the registration environment and the test environment can be prevented. .

3 is a flowchart illustrating a registration method for speaker recognition according to an embodiment. Referring to FIG. 3, the registration device for speaker recognition receives a speaker's voice signal (310). The registration device may receive a predetermined number of voice signals. For example, the registration device may receive the input of the speaker's voice corresponding to the recognition sentence "hi, bixby" repeatedly N times.

The registration device synthesizes a noise signal with the voice signal (320). The noise signal is a signal predetermined to be similar to noise that may be generated in the test step, and may include an additive noise signal, a channel noise signal, or a combination thereof.

The registration device generates a feature vector based on the synthesized signal of the speech signal and the noise signal (330). Various processes may be included to extract feature vectors from the synthesized signal. For example, an image pattern may be generated from a synthesized signal through Fast Fourier Transform (FFT), and a feature vector may be generated by inputting the generated image pattern into an artificial neural network. A method of generating a feature vector will be described in detail with reference to FIGS. 4 to 6 below. Also, the registration device may generate the second feature vector based on the speech signal to which the noise signal is not synthesized.

The registration device constructs a registration database corresponding to the speaker based on the feature vector (340). Also, the registration device may build a registration database based on at least one of a feature vector and a second feature vector. Speaker recognition may be performed in the test step based on the registration database established by steps 310 to 340.

4 is a diagram illustrating a method of generating a feature vector by synthesizing additive noise to a speech signal according to an exemplary embodiment. Referring to FIG. 4, the registration device according to an embodiment may receive a voice signal 410. Since the description of FIGS. 1 to 3 is applicable to FIG. 4 as well, a description of overlapping contents will be omitted.

Additive noise is noise generated from a source different from the voice signal 410 and does not have any correlation between them, and a synthesized signal can be generated by adding the additive noise signal 420 to the voice signal 410. have. When there are a plurality of voice signals 410, synthesized signals may be generated by adding an additive noise signal 420 to each voice signal.

The additive noise signal 420 may be determined in various ways. For example, the additive noise signal 420 may be determined based on at least one of a noise type, a timing of noise, and an energy ratio of noise to a voice signal (SNR). In addition, the additive noise signal 420 may be determined by repetition and/or combination of noise signals determined based on the noise type, the timing of the noise, and the energy ratio (SNR) of noise to the voice signal.

The type of noise may mean the type of noise source such as a babble noise, a door closing sound, and a horn sound. Alternatively, even for noise generated from the same source, the noise type may vary depending on the length of the noise. The timing of noise may mean a time point and/or an end point at which the noise signal is synthesized into the voice signal. The energy ratio (SNR) of noise to the voice signal may mean a relative volume difference between the voice signal and the noise signal.

The registration device may generate a feature vector 450 by inputting a synthesized signal of the speech signal 410 and the additive noise signal 420 to the feature vector generator. The feature vector generator may transform the domain of the synthesized signal and extract a feature vector from the domain transformed result. In relation to the domain conversion, the synthesized signal is a signal generated by adding the speech signal 410 and the additive noise signal 420 and may include sound information in the time domain. The feature vector generator may convert the time domain synthesized signal into a frequency domain form including image information, and generate a feature vector 450 based thereon.

For example, the feature vector generation unit may include a preprocessor 430 and a neural network 440. The preprocessor 430 may convert the time domain synthesized signal into a frequency domain form including image information. For example, the preprocessor 430 may convert the synthesized signal into a frequency domain using a fast Fourier transform to obtain spectrum size information, and based on this, the frequency domain as shown in the drawings 120 to 140 of FIG. Image information in a form can be extracted. The x-axis of the drawings 120 to 140 is in the time domain, the y-axis is in the frequency domain, and the degree of brightness of the drawings may be proportional to the spectral size.

A feature vector 450 may be extracted by inputting a result of domain transformation including image information into the neural network 440. The neural network 440 may be implemented in various structures and may include, for example, a Convolutional Neural Network (CNN) model. The CNN model includes a convolution layer, a pooling layer, and a fully connected layer, and feature vectors are derived from image information input to the model through a computational process performed by each layer. Can be generated. Here, the image information may be pixel values (eg, brightness values) of pixels included in the image.

The above-described preprocessor 430 and neural network 440 correspond to exemplary matters, and the feature vector generator generates feature vectors other than the preprocessor 430 and the neural network 440 structure. It can also be based on other models that can be created. For example, the feature vector generation unit is a linear prediction (LPC) method modeling human sexuality, and a filter bank method modeling a human auditory organ (eg, MFCC based on a Mel-scale filter bank), or It may be based on a combination of these (for example, a method of converting a speech signal based on MFCC and inputting the result to a neural network), or the like.

5 is a diagram for describing a method of generating a feature vector by synthesizing channel noise with a voice signal according to an exemplary embodiment. Referring to FIG. 5, in the case of channel noise, unlike additive noise generated from a source different from the voice signal 510, the channel noise may be noise generated according to a unique characteristic of a place where the voice signal 510 is input. The registration device may generate a synthesized signal by performing a convolution operation of the voice signal 510 and the channel noise signal 520.

The channel noise signal may be determined based on at least one of a noise type and an energy ratio of noise to a voice signal (SNR). For example, the channel noise signal 520 may be determined by repetition and/or combination of noise signals determined based on the type of noise and the energy ratio of noise to voice signal (SNR).

The type of noise may vary depending on the type of place where the ringing occurs, such as a large room, a small room, or a concert hall. Alternatively, the noise type may vary according to microphone characteristics. The energy ratio (SNR) of noise to the voice signal may mean a relative volume difference between the voice signal and the noise signal. Since the description of the feature vector generation unit of FIG. 4 is applicable to FIG. 5 as well, a description of overlapping contents will be omitted.

6 is a diagram illustrating a method of generating a feature vector by segmenting according to an exemplary embodiment. Referring to FIG. 6, the registration device according to an embodiment may receive a voice signal 610.

If the number of utterances of the speaker's speech signal corresponding to the recognition sentence is small in the registration step, it becomes sensitive to the performance of the recognizer for recognizing the starting point of speech, and thus an error is likely to occur in the representative feature vector, which may cause difficulty in recognition. In addition, when a feature vector is generated through a method of using only a specific section of the received voice signal 610 (for example, a center crop method of selecting only the center section of the voice signal), A problem of increasing bias may occur.

The registration device according to an embodiment may segment the voice signal 610 into a plurality of sections 620 to variously register and utilize the limited speech. The segmentation 620 is performed in time series, and the window sizes of each of the sections may be the same. In addition, portions overlapping each other may exist between adjacent sections.

The feature vector generator 630 may generate a feature vector 640 based on a plurality of sections. Specifically, the feature vector generator 630 may extract at least one section feature vector corresponding to at least one of the plurality of sections. The section feature vector may mean a feature vector generated based on the section divided by the segmentation 620. For example, if the registration device segmented one voice signal 610 into M sections 620, M section feature vectors may be generated. Alternatively, if the registration device receives the N voice signals 610 and performs segmentation 620 of each voice signal 610 into M sections, N*M section feature vectors may be generated.

A method of generating a feature vector according to an exemplary embodiment has been described with reference to FIGS. 4 to 6. According to an embodiment, a method of generating a feature vector by synthesizing additive noise to a voice signal described with reference to FIG. 4, a method of generating a feature vector by synthesizing channel noise with a voice signal described with reference to FIG. 5, and A feature vector may be generated in various ways by combining the method of generating a feature vector by segmentation described with reference to FIG. 6.

For example, the registration device may generate N*P*M feature vectors by synthesizing P additive noise signals with N voice signals and segmenting M sections. Alternatively, P1 additive noise signals and P2 channel noise signals may be synthesized with N voice signals, and N*P1*P2*M feature vectors may be generated by segmenting M sections.

The registration device may store the generated feature vector in the registration database to build a registration database corresponding to the speaker for speaker recognition. Based on the established registration database, speaker recognition may be performed in a test step.

7 is a diagram for describing a method of constructing a registration database according to an embodiment. Referring to FIG. 7, the registration device may build a registration database based on a feature vector.

The registration database can be constructed by storing all the generated feature vectors in the registration database, but in this case, it can be affected by outliers a lot, and data processing because it must be compared with all feature vectors in the test stage. Performance may be degraded.

The registration device according to an embodiment clusters a plurality of feature vectors into a plurality of groups, and extracts and registers at least one representative feature vector corresponding to at least one of the plurality of clustered groups. Can build.

The registration device may cluster a plurality of feature vectors into a plurality of groups based on a type of noise or a segmented section. Alternatively, the feature vectors may be clustered into a plurality of groups based on the similarity of the feature vectors. For example, the registration device may cluster the feature vectors according to the k-means algorithm. The k-means algorithm is an algorithm that combines given data into k clusters, and can operate in a manner that minimizes the variance of the difference between each cluster and the distance. However, the clustering method is not limited to the above example. Depending on the design, various methods can be used for clustering.

The registration device may extract at least one representative feature vector corresponding to at least one group from among a plurality of clustered groups. The representative feature vector may be a representative value of the feature vector included in each group. For example, the representative feature vector may be selected from feature vectors included in each group, or may be determined through statistical calculation of feature vectors included in each group. The representative feature vector may be an average value or a median value of feature vectors included in each group. In the case of using a representative feature vector, since the influence of the singular value is small, a robust feature vector can be extracted. The registration device may cluster N*M feature vectors by grouping k, and extract a total of P representative feature vectors by extracting the average value of each group as a representative feature vector. However, the method of extracting the representative feature vector using the representative value is not limited to the above example. The number of feature vectors included in each group may be different, and various methods may be used to extract a representative feature vector according to a design.

8 is a diagram for describing a speaker authentication method according to an embodiment. Hereinafter, the names of the'registered voice signal' 810, the'registered feature vector' 835, and the'representative registered feature vector' 840 are used in the above-described'voice signal', 'Registration' is added to each of'feature vector' and'representative feature vector', and'registered voice signal' 810,'registered feature vector' 835, and'representative registered feature vector' 840 are ' It is substantially the same as'voice signal','feature vector', and'representative feature vector', and the description of FIGS.

In the registration step, the registration device may receive the speaker's registered voice signal 810 and synthesize the additive noise 815 and the channel noise 820. After segmentation 825 of the synthesized signal, the registered feature vector 835 is generated by inputting it to the feature vector generator 830, and a registration database may be constructed by extracting the representative registered feature vector 840. When all of the registered feature vectors 835 are stored in the registration database, the step of extracting the representative registered feature vectors 840 may be omitted.

The recognition device for speaker recognition may receive an input voice signal 845 of a speaker. The input voice signal 845 may mean a voice signal according to a speaker's speech for authentication in a test step. Since the test step is performed in a noisy environment, the input speech signal 845 may naturally contain noise.

A recognition device for speaker recognition is a device that determines whether or not to recognize (including authentication) by generating a feature vector from the speaker's input voice and comparing it with the feature vector stored in the registration database. For example, a software module, a hardware module, or these It can be implemented in a combination of. The recognition device may be mounted on various computing devices and/or systems such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a television, a wearable device, a security system, and a smart home system.

The recognition device according to an embodiment may include a registration device, and in this case, the recognition device performing the test step may be referred to as a'consensus recognition device', and a recognition device including a registration device may be referred to as a'wide recognition device. It can be said. The distinction between the recognition device and the registration device is to distinguish and describe each function. Therefore, when a product is actually implemented, it can be configured to process all of them in one device ('broad sense recognition device'). For example, the acoustic sensor of the registration device that receives the registered voice signal 810 and the acoustic sensor of the'negotiation recognition device' that receives the input voice signal 845 may be the same. In addition, the feature vector generator 830 of the registration device and the feature vector generator 855 of the'negotiation recognition device' may be the same. Hereinafter, for convenience of description, the recognition device will be described as meaning a'negotiate recognition device'.

The recognition apparatus may divide and extract the speaker's recognition speech signal 845 so that the window size of the input feature vector is the same as the window size of the registered feature vector. For example, the recognition apparatus may center crop the speaker's recognized speech signal 845 or perform a sliding window 850 on the speaker's recognized speech signal 845. This is because each vector can be compared when the window size of the input feature vector is the same as the window size of the registered feature vector. Hereinafter, for convenience of explanation, the sliding window 850 is used as a reference, but various methods of dividing and extracting the recognized speech signal 845 may be applied in embodiments.

The recognition device may generate an input feature vector 860 by inputting the result of performing the sliding window 850 into the feature vector generator 855. As described above, the feature vector generator 855 may generate the input feature vector 860 using the same neural network as the feature vector generator 830.

The recognition device may compare (865) at least one registered feature vector 835 or representative registered feature vector 840 of the registered user stored in the registration database established in the registration step with at least one input feature vector 860 . The registered user may mean a speaker corresponding to the registered feature vector (or representative registered feature vector) stored in the registration database.

The recognition device may authenticate 870 a speaker based on the result of the comparison 865. The recognition device may authenticate the speaker 870 based on the registered feature vector 835 or the similarity score between the representative registered feature vector 840 and the input feature vector 860. For example, when the representative value of the similarity score is equal to or greater than a predetermined threshold, the recognition device may determine that the speaker authentication 870 is successful. Alternatively, the recognition device may determine that the speaker authentication 870 has succeeded only when the number of registered feature vectors 835 or representative registered feature vectors 840 having a similarity score with the input feature vector 860 equal to or greater than a threshold is greater than a predetermined number. May be. However, the condition of the speaker authentication 870 is not necessarily limited to the above example, and various conditions may be applied.

The similarity score may be determined in the form of, for example, a matching score or a distance score. The matching score may correspond to a degree of similarity between the registered feature vector 835 or the representative registered feature vector 840 and the input feature vector 860, that is, a score indicating the degree of similarity between feature vectors. A low matching score may mean that the similarity between the feature vectors is low, and a high match score may mean that the similarity between the feature vectors is high. In addition, the distance score may correspond to a score representing a feature distance (e.g., Euclidean distance) between the registered feature vector 835 or the representative registered feature vector 840 and the input feature vector 860. have. A low distance score means that a feature distance between feature vectors is close in a feature vector space, and a high distance score means that feature distances between feature vectors are far.

9 is a flowchart illustrating a speaker recognition method according to an exemplary embodiment. Referring to FIG. 9, the recognition device receives a speaker's voice signal (910).

The recognition device generates at least one input feature vector based on the speech signal (920). The number of input feature vectors may be determined based on the security level or response time required by the application. For example, a speaker authentication method according to an embodiment may be applied to a payment application or to unlock a smart phone. When a speaker authentication method is applied to unlock a smartphone, user convenience is more important than when used in a payment application, so a smaller number of input feature vectors may be sufficient than when used in a payment application. .

The recognition apparatus may domain-transform the speaker's speech signal and extract an input feature vector from the domain-transformed result. Specifically, the recognition apparatus may perform fast Fourier transform on the speaker's speech signal, and may perform padding on the result of fast Fourier transform according to the result of fast Fourier transform.

There may be no data of a specific section in the result of performing the fast Fourier transform. In this case, the data can be filled through padding. As a result of the fast Fourier transform, when there is no data of a specific section, padding may be performed based on data of an adjacent section. This can be based on the assumption that the voice signal is a continuous signal and that the data value does not change rapidly for a sufficiently short time. For example, if the section in which data is lost includes a plurality of unit sections, the data of the first lost section is determined to be the same value as the data of the immediately adjacent unlost unit section, and the data of the second lost section is the next adjacent unit section. It can be determined by the same value as the data of.

The recognition device compares at least one registered feature vector of the registered user stored in the registration database with at least one input feature vector (930), and recognizes the speaker based on the comparison result (940). Speaker recognition may include speaker authentication and speaker identification. Speaker authentication is to certify whether the speaker is the person or not, and speaker identification can mean finding out who the speaker is among several registered speakers. In the following, a speaker identification method will be described with reference to FIG. 10.

10 is a diagram for describing a speaker identification method according to an exemplary embodiment. Referring to FIG. 10, a method of identifying a speaker according to an embodiment may relate to a method of finding out who the speaker is among several registered speakers. The speaker identification method is a case in which there are a plurality of registered users, and can be operated by performing speaker authentication multiple times.

The registration step of the speaker identification method may be performed in the same manner as the registration step of the speaker authentication method described with reference to FIG. 8. However, in the case of the speaker authentication method, a feature vector for one speaker is generated, but in the case of the speaker identification method, a feature vector for a plurality of speakers, for example, N speakers, is generated, and based on this, N speakers are registered. Build a database.

The test step of the speaker identification method may also be performed in the same manner as the test step of the speaker authentication method described with reference to FIG. 8. For example, by measuring the similarity between the registration feature vector 1020 generated for each individual or the representative registration feature vector 1025 generated for each individual and the input feature vector 1040 (1045), the similarity numerical vector 1050 Can be determined. The similarity numerical vector 1050 may be an example of a similarity score. According to the speaker identification method, a registered feature vector 1020 having the highest similarity value vector 1050 with the input feature vector 1040 or a registered user corresponding to the representative registered feature vector 1025 is identified as a speaker (1055 )can do. However, even in this case, the size of the similarity numerical vector must be greater than or equal to the threshold. If the sizes of all the similarity numerical vectors 1050 are less than or equal to the threshold, it may be output that they do not correspond to registered users. Alternatively, a user with the largest number of registered feature vectors 1020 or representative registered feature vectors 1025 having a size of the similarity numerical vector 1050 greater than or equal to a threshold may be identified as a speaker (1055). Alternatively, a registered feature vector 1020 having a size of the similarity numerical vector 1050 greater than or equal to a threshold value or a user having the largest representative value of the representative registered feature vector 1025 may be identified as a speaker (1055). However, the condition of the speaker identification 1055 is not necessarily limited to the above example, and various conditions may be applied.

Fig. 11 is a block diagram of a registration device for speaker recognition according to an embodiment. Referring to FIG. 11, the registration device 1100 according to an embodiment includes a processor 1110. The registration device 1100 may further include a memory 1130, a communication interface 1150, and sensors 1170. The processor 1110, the memory 1130, the communication interface 1150, and the sensors 1170 may communicate with each other through the communication bus 1105.

The processor 1110 synthesizes a noise signal with a voice signal, generates a feature vector based on the synthesized signal of the voice signal and the noise signal, and builds a registration database corresponding to the speaker based on the feature vector.

The memory 1130 may include a registration database including feature vectors. The memory 1130 may be a volatile memory or a non-volatile memory.

The sensors 1170 may include, for example, an acoustic sensor that receives a speaker's voice signal. Sensors can collect various acoustic information.

According to an embodiment, the processor 1110 may generate a second feature vector based on a voice signal, and build a registration database based on at least one of the feature vector and the second feature vector. The processor 1110 may domain transform the synthesized signal and extract a feature vector from the domain transformed result. The processor 1110 may add an additive noise signal to the voice signal. The processor 1110 may perform a convolution operation of a voice signal and a channel noise signal. The processor 1110 may segment the synthesized signal of the speech signal and the noise signal into a plurality of sections, and extract at least one section feature vector corresponding to at least one of the plurality of sections. The processor 1110 clusters a plurality of feature vectors including at least one of a feature vector and a second feature vector into a plurality of groups, and generates at least one representative feature vector corresponding to at least one of the plurality of groups. Can be extracted.

In addition, the processor 1110 may perform at least one method or an algorithm corresponding to at least one method described above with reference to FIGS. 2 to 7. The processor 1110 may execute a program and control the registration device 1100. Program codes executed by the processor 1110 may be stored in the memory 1130. The registration device 1100 is connected to an external device (eg, a personal computer or a network) through an input/output device (not shown), and can exchange data. The registration device 1100 may be mounted on various computing devices and/or systems such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a television, a wearable device, a security system, and a smart home system.

The recognition device according to an embodiment includes a processor. The recognition device may further include a memory, a communication interface, and sensors. The processor, memory, communication interface, and sensors may communicate with each other through a communication bus.

The processor generates at least one input feature vector based on the speech signal, compares at least one registered feature vector of the registered user with at least one input feature vector stored in the registration database, and recognizes a speaker based on the comparison result. .

The memory may include a registration database containing feature vectors. The registration database may include at least one registration feature vector generated based on a synthesized signal of a voice signal and a noise signal for registration of a registered user. The memory may be a volatile memory or a non-volatile memory.

The sensors may include, for example, an acoustic sensor for receiving a speaker's voice signal. Sensors can collect various acoustic information.

According to an embodiment, the processor may perform domain transformation on the speaker's speech signal and extract an input feature vector from the domain transformation result. The processor may perform a sliding window on the speaker's speech signal so that the window size of the input feature vector is the same as the window size of the registered feature vector. The processor may perform a fast Fourier transform on a speaker's speech signal, and may perform padding on a result of the fast Fourier transform according to the result of the fast Fourier transform.

The recognition device 1100 may be mounted on various computing devices and/or systems such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a television, a wearable device, a security system, and a smart home system.

12 is a diagram illustrating a method of authenticating a speaker through a plurality of channels according to an exemplary embodiment. Referring to FIG. 12, the registration step and the test step according to an embodiment may be performed through a plurality of channels.

The registration device may include a plurality of acoustic sensors, and a channel may be configured for each acoustic sensor. A channel means a path through which information or signal transmission is guaranteed, and since each channel is independent, signals of each channel can have independence without interference. Hereinafter, for convenience of description, the description is based on an environment including two channels, but the embodiments may be equally applied to an environment including three or more channels.

In the registration step, each of the acoustic sensors of the registration device may receive the speaker's registered voice signal 1 (1210) and the speaker's registered voice signal 2 (1211). The speaker's registered voice signal 1 (1210) and the speaker's registered voice signal 2 (1211) are signals generated simultaneously from the same sound source.For example, the speaker's registered voice signal 1 (1210) is a microphone input located at the bottom of the registration device. The registered signal, and the registered voice signal 2 1211 may be a signal input to a microphone located at the top of the registration device.

The operation of the registration step through a plurality of channels is only performed independently through a plurality of channels, and the specific method of the registration step through a single channel described with reference to FIG. 8 can be applied equally to the registration step through a plurality of channels. have. For example, after synthesizing the additive noise 1215 and the channel noise 1220 to the speaker's registered voice signal 1 1210 received by the microphone located at the bottom of the registration device, and segmenting the synthesized signal 1225, A registered feature vector can be generated by inputting it into the feature vector generator 1230, and the registered voice signal 2 1211 of the speaker received by the microphone located at the top of the registration device also independently generates a registered feature vector through the same method. can do.

A representative registration feature vector 1240 may be extracted based on the registration feature vector generated in each channel, and a registration database may be constructed based on this. Registration feature vectors may be diversified by generating a registration feature vector in each of a plurality of channels.

Like the registration step, the test step may be independently performed through a plurality of channels. For example, the speaker's input voice signal 1 (1245) is input to a microphone located at the bottom of the registration device, and the registered voice signal signal 2 (1246) is input to a microphone located at the top of the registration device, and input feature vectors independently from each channel Can be created. Likewise, since input feature vectors are generated in each of a plurality of channels, input feature vectors may be diversified.

13 is a diagram for explaining a speaker authentication method through channel selection according to an embodiment.

Referring to FIG. 13, the registration device according to an embodiment may include a plurality of acoustic sensors, and a channel may be configured for each acoustic sensor. Hereinafter, for convenience of description, the description is based on an environment including two channels, but the embodiments may be equally applied to an environment including three or more channels. The description of FIGS. 1 to 12 is applicable to FIG. 13 as well. Description of the duplicated content will be omitted.

Each of the acoustic sensors of the registration device according to an embodiment may receive a speaker's voice signal 1 (1310) and a speaker's registered voice signal 2 (1311). The registration device may pre-process (1320) the received speaker's voice signals 1310 and 1311.

For example, the registration device may estimate noise based on the difference between the speaker's voice signals 1310 and 1311, and subtract the noise from each of the speaker's voice signals 1310 and 1311 to obtain a signal from which noise has been removed. Can be obtained. For example, a common portion of the speaker's voice signals 1310 and 1311 may have a high probability of corresponding to the speaker's voice, and a difference between the speaker's voice signals 1310 and 1311 may have a high probability of corresponding to noise. By using this, noise can be estimated based on the difference between the speaker's speech signals 1310 and 1311, and noise is removed by subtracting the noise from each of the speaker's speech signals 1310 and 1311. I can.

As another example, the registration device may select any one of the speaker's voice signals received through a plurality of channels. When a voice signal of more than an allowable input size is input to an acoustic sensor of a registration device in which an allowable limit input exists, a recognition error may occur due to a hardware specification of the registration device such as clipping. For example, clipping may occur when a voice signal is input from a location too close to an acoustic sensor or a voice signal is input with too loud a voice. According to the speaker recognition method through a single channel, recognition errors may occur due to clipping. The registration device may reduce such an error by selecting any one of the speaker's voice signals received through a plurality of channels for each section.

For example, when a voice signal is input from a location close to the acoustic sensor at the bottom of the registration device, the speaker's voice signal 1310 received from the acoustic sensor at the bottom of the registration device may be clipped, but from the acoustic sensor at the top of the registration device. The received speaker's voice signal 1311 may be normally input. In this case, it is possible to solve the clipping problem by selecting the speaker's voice signal 1311 received through the acoustic sensor on the top of the registration device in which clipping has not occurred in the preprocessing step 1320.

Furthermore, the registration device may select any one of the speaker's voice signals received through a plurality of channels for each section. For example, in the first section, the audio signal 1 (1310) may be selected, and in the second section, the audio signal 2 (1311) may be selected.

Subsequent operations 1330 to 1380 may be applied in the same manner as operations 220 to 270 of the registration system for speaker recognition described with reference to FIG. 2.

14 is a diagram for describing a method of overcoming a change in vocalization according to a location and direction of a registration device according to an exemplary embodiment.

When the registration device receives a voice signal through a single channel, a recognition error may occur due to a change in vocalization according to the echo characteristics of the registration device. For example, in the registration step 1400, the acoustic sensor of the registration device is located in an open space so that there is no reverberation by the table, and in the test step 1450, the acoustic sensor is located on the table and there may be reverberation by the table. have. In this case, a recognition error may occur due to a change in vocalization according to the echo characteristics of the registration device.

When the registration device adds a registration feature vector using a plurality of channels, it is possible to increase a speaker recognition rate by covering a change in speech according to the direction of the registration device. For example, in the registration step 1400, the acoustic sensor at the bottom of the registration device is located in an open space, but the acoustic sensor at the top of the registration device is located on the table, so that registration can be performed in the same environment as the test step 1450. . Accordingly, when the registration device uses a plurality of channels, it is possible to cover a change in speech according to the echo characteristics of the registration device.

In addition, even if the registration device uses a plurality of channels, in an environment in which a plurality of channels can be processed in parallel, for example, in a multithreading environment, the speaker recognition rate can be increased without causing additional delay time during speaker recognition. I can.

The embodiments described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices, methods, and components described in the embodiments include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA). array), programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions, such as one or more general purpose computers or special purpose computers. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to behave as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodyed in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of the program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

As described above, although the embodiments have been described by the limited drawings, a person of ordinary skill in the art can apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order from the described method, and/or components such as a system, structure, device, circuit, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and claims and equivalents fall within the scope of the claims to be described later.

Claims (38)

In the registration method for speaker recognition,
Receiving a speaker's voice signal;
Synthesizing a noise signal to the speech signal;
Generating a feature vector based on a synthesized signal of the speech signal and the noise signal; And
Constructing a registration database corresponding to the speaker based on the feature vector
Registration method comprising a.
The method of claim 1,
Generating a second feature vector based on the speech signal
Including more,
The step of building the registration database
Building the registration database based on at least one of the feature vector and the second feature vector
Including, registration method.
The method of claim 1,
Generating the feature vector comprises:
Domain transforming the synthesized signal; And
Extracting the feature vector from the domain transformed result
Including, registration method.
The method of claim 1,
The synthesizing step
Adding an additive noise signal to the speech signal
Including, registration method.
The method of claim 1,
The synthesizing step
Performing a convolution operation of the voice signal and channel noise signal
Including, registration method.
The method of claim 4,
The additive noise signal is
The registration method, wherein the registration method is determined based on at least one of a type of noise, a timing of noise, and an energy ratio of noise to the audio signal (SNR).
The method of claim 5,
The channel noise signal is
The registration method is determined based on at least one of a type of noise and an energy ratio of noise to the speech signal (SNR).
The method of claim 1,
Generating the feature vector comprises:
Segmenting the synthesized signal of the speech signal and the noise signal into a plurality of sections; And
Extracting at least one section feature vector corresponding to at least one of the plurality of sections
Including, registration method.
The method of claim 2,
The step of building the registration database
Clustering a plurality of feature vectors including at least one of the feature vector and the second feature vector into a plurality of groups;
Extracting at least one representative feature vector corresponding to at least one of the plurality of groups
The registration method further comprising.
The method of claim 1,
Extracting a start point and an end point of sound information included in at least one of the voice signal and the synthesized signal
Including more,
The registration method, wherein the starting point and the ending point are applied to at least one of the synthesizing and generating.
The method of claim 3,
The synthesized signal includes sound information in a time domain, and the domain transformed result includes image information in a frequency domain,
Generating the feature vector comprises:
Extracting the feature vector from image information in the frequency domain using a convolution neural network (CNN)
Including, registration method.
The method of claim 1,
In the case of receiving the speaker's voice signal through a plurality of channels in the receiving step,
The synthesizing and generating are independently performed corresponding to each of the plurality of channels,
The registration method, wherein the registration database is built based on feature vectors generated by the plurality of channels in the building step.
The method of claim 12,
After the receiving step,
Estimating noise based on differences between the speaker's voice signals received through the plurality of channels; And
Subtracting the noise from each of the speaker's voice signals
The registration method further comprising.
The method of claim 1,
Receiving the speaker's voice signal comprises:
Receiving the speaker's voice signals through a plurality of channels,
Determining the voice signal for the synthesizing step by preprocessing the voice signals of the speaker received through the plurality of channels
The registration method further comprising.
The method of claim 14,
The determining step
Selecting one of the voice signals based on whether a recognition error due to a hardware spec is included in the speaker's voice signals received through the plurality of channels;
Selecting any one of the voice signals for each section based on whether the recognition error is included for each section of the speaker's voice signals received through the plurality of channels
Including, registration method.
Receiving a speaker's voice signal;
Generating at least one input feature vector based on the speech signal;
Comparing at least one registered feature vector of a registered user stored in a registration database with the at least one input feature vector; And
Recognizing the speaker based on the comparison result
Including,
The registration database includes the at least one registration feature vector generated based on a synthesized signal of a voice signal and a noise signal for registration of the registered user.
The method of claim 16,
When there are multiple registered users
Recognizing the speaker is
Including the step of identifying the speaker, speaker recognition method.
The method of claim 16,
Generating the input feature vector comprises:
Domain-converting the speaker's voice signal; And
Extracting the input feature vector from the domain transformed result
Including, speaker recognition method.
The method of claim 16,
Generating the input feature vector comprises:
Performing a sliding window on the speaker's speech signal so that the window size of the input feature vector is the same as the window size of the registered feature vector.
Including, speaker recognition method.
The method of claim 16,
The number of input feature vectors is determined based on a security level or response time required by an application.
The method of claim 18,
The steps to convert the domain are
Performing a fast Fourier transform on the speaker's speech signal; And
Performing padding on the result of the fast Fourier transform according to the result of the fast Fourier transform
Including, speaker recognition method.
The method of claim 16,
The noise signal is
Including at least one of additive noise (additive noise) signal and channel noise (channel noise) signal, speaker authentication method.
The method of claim 16,
In the case of receiving the speaker's voice signal through a plurality of channels in the receiving step,
The generating step and the comparing step are independently performed corresponding to each of the plurality of channels,
The registration feature vector includes registration feature vectors generated by the plurality of channels.
The method of claim 23,
After the receiving step,
Estimating noise based on differences between the speaker's voice signals received through the plurality of channels; And
Subtracting the noise from each of the speaker's voice signals
Further comprising a, speaker authentication method.
The method of claim 1,
Receiving the speaker's voice signal comprises:
Receiving the speaker's voice signals through a plurality of channels,
Determining the voice signal for the generating step by preprocessing the speaker's voice signals received through the plurality of channels
Further comprising a, speaker authentication method.
The method of claim 25,
The determining step
Selecting one of the voice signals based on whether a recognition error due to a hardware spec is included in the speaker's voice signals received through the plurality of channels;
Selecting any one of the voice signals for each section based on whether the recognition error is included for each section of the speaker's voice signals received through the plurality of channels
Including, speaker authentication method.
A computer program combined with hardware and stored in a medium for executing the method of any one of claims 1 to 26.
In the registration device for speaker recognition,
An acoustic sensor for receiving a speaker's voice signal;
Processor for synthesizing a noise signal with the speech signal, generating a feature vector based on the synthesized signal of the speech signal and the noise signal, and constructing a registration database corresponding to the speaker based on the feature vector
Registration device comprising a.
The method of claim 28,
The processor is
And generating a second feature vector based on the speech signal, and constructing the registration database based on at least one of the feature vector and the second feature vector.
The method of claim 28,
The processor is
Domain transforming the synthesized signal, and extracting the feature vector from the domain transformed result.
The method of claim 28,
The processor is
A registration device for adding an additive noise signal to the speech signal.
The method of claim 28,
The processor is
A registration device that performs a convolution operation of the speech signal and a channel noise signal.
The method of claim 28,
The processor is
A registration device for segmenting the synthesized signal of the speech signal and the noise signal into a plurality of sections, and extracting at least one section feature vector corresponding to at least one of the plurality of sections.
The method of claim 28,
The processor is
Clustering a plurality of feature vectors including at least one of the feature vector and the second feature vector into a plurality of groups, and extracting at least one representative feature vector corresponding to at least one of the plurality of groups , Registration device.
An acoustic sensor for receiving a speaker's voice signal; And
Generate at least one input feature vector based on the voice signal, compare at least one registered feature vector of a registered user stored in a registration database with the at least one input feature vector, and select the speaker based on the comparison result. Including a processor that recognizes,
The registration database includes the at least one registration feature vector generated based on a synthesized signal of a voice signal and a noise signal for registration of the registered user.
The method of claim 35,
The processor is
The speaker recognition apparatus for performing domain conversion of the speaker's speech signal and extracting the input feature vector from the domain conversion result.
The method of claim 35,
The processor is
A speaker recognition apparatus for performing a sliding window on the speaker's speech signal so that the window size of the input feature vector is the same as the window size of the registered feature vector.
The method of claim 35,
The processor is
A speaker recognition apparatus for performing a fast Fourier transform on the speech signal of the speaker, and performing padding on a result of the fast Fourier transform according to a result of the fast Fourier transform.

Images