KR101304127B1 - Apparatus and method for recognizing of speaker using vocal signal - Google Patents

Abstract

The present invention relates to a speaker recognition apparatus using a voice signal and a method thereof, and a speaker recognition apparatus using a voice signal according to an embodiment of the present invention, a voice receiving unit for receiving a voice signal, and the received voice signal frequency Selects M subsets that are higher than a preset recognition rate by applying a recognition rate weight to the N subsets included in the extracted feature and the feature extracting unit for converting to the region and extracting the feature from the converted speech signal And a speaker matching unit configured to calculate a surrounding likelihood score and recognize a speaker that has generated the voice signal based on a maximum likelihood score which is the largest value among the surrounding likelihood scores.
Accordingly, by using the loss characteristic theory for speech recognition using the speech signal, the characteristics of the speech signal are extracted and only the subset of the characteristic having the highest recognition rate is calculated to reduce the calculation amount and increase the recognition accuracy.

Description

Speaker recognition apparatus using voice signal and its method {APPARATUS AND METHOD FOR RECOGNIZING OF SPEAKER USING VOCAL SIGNAL}

The present invention relates to a speaker recognition apparatus using a speech signal and a method thereof, and more particularly, a speaker recognition technique using loss characteristic theory is disclosed.

Biometrics are increasing in importance and utility for various security devices used by modern people for security reasons. Among the fields of biometric technology, speaker recognition technology is particularly useful because it can be implemented as a simple interface between a person and a computer. However, the performance of such speaker recognition technology is degraded by the influence of background noise. Therefore, in recent years, research on powerful speaker recognition technology has been conducted to minimize the influence of background noise and to reduce background noise.

The conventional speaker recognition system extracts characteristics by Mel Frequency Cepstral Coefficient (MFCC) using an auditory model. Then, a model of each speaker is generated using a Gaussian mixture model (GMM), and the speaker is recognized for the input voice. However, there is a problem that the speaker recognition success rate is very low for a voice having a heavy background noise.

Meanwhile, in another conventional speaker recognition system, a Missing Feature Theory (MFT) extracts a plurality of features from a speech signal to recognize a speaker, and a portion of the extracted features that is deteriorated very much by noise. Defined as unreliable, only trusted properties are used, except untrusted properties. To select reliable characteristics, prior knowledge of noise statistics can be used.

On the other hand, the Extended Missing Feature Theory (EMFT) does not use any prior knowledge of noise statistics. The Extended Loss Characteristic Theory (EMFT) ignores changes in noise outside the statistics of the training database to reduce the error between the training database and the test database. Since noise statistics vary over time or are not well known, extended loss characteristic theory (EMFT) can be a good way to implement speaker recognition techniques for noise mixed speech signals that change rapidly over time. .

However, in loss characteristic theory (MFT) and extended loss characteristic theory (EMFT), likelihood values are calculated for all possible combinations of reliable features among extracted features, and have a maximum likelihood value. Finds a combination of characteristics and uses it for speaker recognition. Thus, to calculate the likelihood values of all possible combinations of reliable properties, there has been a problem that is very complicated and requires a large amount of calculation.

The speaker recognition system using the existing MFCC and the GMM is much vulnerable to the background noise. The MMF and the EMFT are stronger in the background noise than the speaker recognition system using the GMM, but require a large amount of computation. Accordingly, the present inventors have studied a speaker recognition system that is robust against background noise and has a low computational amount.

The background technology of the present invention is described in Republic of Korea Patent Publication No. 10-1060162 (2011. 08. 23).

An object of the present invention is to reduce the amount of calculation and increase the recognition accuracy while using the loss characteristic theory for speaker recognition using speech signals.

A speaker recognition apparatus using a voice signal according to an embodiment of the present invention, a voice receiver for receiving a voice signal, and a feature extraction for converting the received voice signal into a frequency domain and extracting a characteristic from the converted voice signal A neighboring likelihood score is calculated by applying a recognition rate weight to the N subsets included in the extracted feature, selecting M subsets higher than a preset recognition rate, and calculating a maximum likelihood score, which is the largest value among the surrounding likelihood scores. And a speaker matching unit configured to recognize the speaker who has generated the voice signal based on the likelihood score.

The speaker matching unit may further include: a voice signal DB for storing voice signal data of at least one or more speakers, and M portions having a recognition rate weight applied to N subsets included in the extracted feature to be higher than a preset recognition rate. A subset selector for selecting a set, a peripheral likelihood score calculator for calculating a surrounding likelihood score for the received speech signal using the selected M subsets, and a maximum likelihood score which is the largest value among the peripheral likelihood scores; Based on the voice signal data of all the speakers stored in the voice signal DB based on the speaker recognition unit for recognizing the speaker may be included.

The apparatus may further include a device controller configured to control an operation of a networked peripheral device in response to the speaker recognition result.

)를 다음의 수학식을 이용하여 계산된 값을 정규화하고, 상기 인식률 가중치가 기 설정된 인식률보다 높은 M 개의 부분 집합을 선택할 수 있다:The subset selection unit may include the recognition rate weight (

) Can be normalized using the following equation, and M subsets can be selected in which the recognition rate weight is higher than a preset recognition rate:

Here, w _N is a weight when the subset is N, R _N is a speaker recognition rate, U _N is a speaker recognition rate.

The peripheral likelihood score calculator may obtain the peripheral likelihood score P (λ _S | X _M ) using the following equation:

Here, lambda _S represents the model of the speaker S, and X _M represents the speaker's characteristic vector in the case of M subsets.

The speaker recognition unit may obtain the maximum likelihood score P (X | λ _S ) using the following equation:

은 인식률 가중치를 나타낸다.Where λ _S is the speaker S model, X _M is the speaker's characteristic vector in the case of M subsets,

Denotes a recognition rate weight.

A speaker recognition method for allowing a speaker to access a device by recognizing a speaker who has generated a voice signal,

In another embodiment, a speaker recognition method using a speech signal includes receiving the speech signal, converting the received speech signal into a frequency domain, and extracting a characteristic from the converted speech signal. And applying a recognition rate weight to the N subsets included in the extracted feature, selecting M subsets higher than a preset recognition rate to calculate a neighbor likelihood score, and calculating a maximum likelihood value that is the largest value among the neighbor likelihood scores. Matching the speaker that generated the speech signal based on a score.

As described above, according to the present invention, by using loss characteristic theory for speech recognition using a speech signal, the speech signal may be extracted and calculated only for a subset of the characteristic having the highest recognition rate, thereby reducing the amount of calculation and increasing the recognition accuracy. .

1 is a block diagram of a speaker recognition apparatus using a voice signal according to an embodiment of the present invention,
2 is a detailed configuration diagram of a speaker matching unit in the speaker recognition apparatus of FIG. 1;
3 is a flowchart of a speaker recognition method implemented through the speaker recognition apparatus of FIG. 1;
4A to 4C are exemplary diagrams for describing setting a recognition rate weight according to the speaker recognition apparatus of FIG. 1.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The terms used are terms selected in consideration of the functions in the embodiments, and the meaning of the terms may vary depending on the user, the intention or the precedent of the operator, and the like. Therefore, the meaning of the terms used in the following embodiments is defined according to the definition when specifically defined in this specification, and unless otherwise defined, it should be interpreted in a sense generally recognized by those skilled in the art.

1 is a block diagram of a speaker recognition apparatus using a voice signal according to an embodiment of the present invention, Figure 3 is a flowchart of a speaker recognition method implemented by the speaker recognition apparatus according to FIG.

1 and 3, the speaker recognition apparatus 100 using the voice signal according to an embodiment of the present invention includes a voice receiver 110, a feature extractor 120, and a speaker matcher 130. . The voice receiver 110 receives a voice signal from a speaker (S300). In this case, the voice signal received by the voice receiver 110 may be a voice signal in which various background noises are mixed with the original voice signal generated by the speaker. For example, a voice signal generated by a speaker may be mixed with a car sound having various frequencies and sizes, a mobile phone ringing sound, a conversation sound of people around you, an airplane sound, and the like. The voice receiver 110 outputs the input voice input to the feature extractor 120.

Next, the feature extractor 120 converts the voice signal received from the voice receiver 110 into a frequency domain and extracts a feature from the converted voice signal (S310). Even if the voice signal is the same language, it is not only very complicated depending on the gender, age, state of the pronunciation of the speaker, but also changes its properties when pronounced alone and whenever it is pronounced in a word or sentence. Feature extraction that can express features is important. In other words, it is necessary to extract information that can increase the coherence between the same voice signals and at the same time distinguish them from other voice signals. This information is called a feature of the voice signal.

The feature extractor 120 may extract a plurality of speech features to be used for speaker recognition using a filter or a window in a specific frequency region of the input speech signal. For example, the feature extractor 120 may use techniques such as mel-frequency cepstral coefficients (MFCC), linear prediction coefficients (LPC), cepstrum, and energy for each frequency band. A plurality of characteristics may be extracted from the input voice signal.

Next, the speaker matching unit 130 applies a recognition rate weight to the N subsets included in the feature extracted from the feature extractor 120, and selects M subsets having a higher recognition rate than the preset recognition rate to obtain a surrounding likelihood score. The marginal likelihood score is calculated and the speaker who generates the speech signal is recognized based on the maximum likelihood score, which is the largest value among the surrounding likelihood scores (S320). That is, the speaker corresponding to the voice signal data having the largest maximum likelihood score is recognized as the speaker who generated the voice signal. The detailed configuration and function of the speaker matching unit 130 will be described later in detail with reference to FIG. 2.

On the other hand, the speaker recognition apparatus 100 according to another embodiment of the present invention may further include a device controller 140. The device controller 140 controls the operation of the networked peripheral device in response to the speaker recognition result. For example, the device controller 140 receives information about a speaker who has generated a voice signal from the speaker matching unit 130, and accesses and golds in a building that is a peripheral device networked with the speaker recognition device 100 only to the speaker. Access to various devices, such as torture and the use of mobile terminals, may be allowed.

FIG. 2 is a detailed configuration diagram of a speaker matching unit in the speaker recognition apparatus of FIG. 1.

Referring to FIG. 2, the speaker matching unit 230 includes a subset selector 231, a peripheral likelihood score calculator 232, a speaker recognizer 233, and a voice signal DB 234. The subset selector 231 selects M subsets higher than a preset recognition rate by applying a recognition rate weight to the N subsets included in the extracted feature. Here, the recognition rate weight refers to a weight given to a subset that greatly affects speaker recognition through pre-learned data about the speaker recognition and the speaker recognition among the N subsets.

)를 다음의 수학식 1을 이용하여 계산된 값을 최소-최대값에 대해 정규화하고, 인식률 가중치가 기 설정된 인식률보다 높은 M 개의 부분 집합을 선택할 수 있다.In addition, the subset selection unit 231 may recognize the recognition rate weight (

) Can be normalized to the minimum-maximum value using Equation 1 below, and M subsets having a recognition rate weight higher than a predetermined recognition rate can be selected.

=

(when

>δ_TH, others 0)를 이용하여, 기 설정된 인식률 문턱치(δ_TH) 보다 큰 경우 외에는 값을 0으로 설정하는 것도 가능하다.In Equation 1, w _N denotes a weight when the subset is N, R _N denotes a speaker recognition rate, and U _N denotes a speaker recognition rate. In addition, the recognition rate weight mask to reduce the computational complexity

=

(when

> δ _TH , others 0), it is also possible to set the value to 0 except when larger than the preset recognition rate threshold δ _TH .

Setting the recognition rate weight will be described later with reference to FIGS. 4A to 4C.

4A to 4C are exemplary diagrams for describing setting a recognition rate weight according to the speaker recognition apparatus of FIG. 1.

)를 나타낸다. 그 결과, 음성 신호의 특성의 부분 집합 개수 N이 8, 9인 경우에 화자 인식 성능이 가장 높은 것으로 나타났다.4A to 4C show a result of applying an advanced missing feature theory (AMFT) using only a vowel of voice signals. 4A shows a recognition rate R when a speaker is recognized, and FIG. 4B shows a false recognition rate U when a speaker is not recognized. FIG. 4C illustrates the recognition rate weight calculated by substituting into Equation 1 described above and normalized using the results of FIGS. 4A and 4B.

). As a result, the speaker recognition performance was found to be highest when the number of subsets N of the characteristics of the speech signal was 8 or 9.

)를 기준으로 인식률이 높은 부분 집합 개수 N을 변경할 수 있다. 예를 들어, 기 설정된 인식률의 문턱치가 0.3 보다 높은 경우는 N이 {2, 6, 7, 8, 9 10}인 경우가 선택되며, 인식률 문턱치가 0.45 보다 높은 경우는 N이 {6, 7, 8, 9 10}인 경우가 선택되며, 인식률 문턱치가 0.7 보다 높은 경우는 N이 {7, 8, 9 10}인 경우가 선택되며, 인식률 문턱치가 0.9 보다 높은 경우는 N이 {8, 9}인 경우가 선택되며, 인식률 문턱치가 0.98 보다 높은 경우는 N이 {9}인 경우가 선택된다.
In addition, in FIG. 4C, the recognition rate weight (

), The number of subsets N having a high recognition rate can be changed. For example, when the threshold of the preset recognition rate is higher than 0.3, the case where N is {2, 6, 7, 8, 9 10} is selected. When the recognition rate threshold is higher than 0.45, N is {6, 7, 8, 9 10} is selected. When the recognition rate threshold is higher than 0.7, N is selected as {7, 8, 9 10}. When the recognition rate threshold is higher than 0.9, N is {8, 9}. Is selected. When the recognition rate threshold is higher than 0.98, the case where N is {9} is selected.

Referring back to FIG. 2, the surrounding likelihood score calculator 232 calculates the surrounding likelihood score for the received speech signal using the M subsets selected by the subset selecting unit 231. For example, the peripheral likelihood score calculator 232 can obtain the peripheral likelihood score P (λ _S | X _M ) using the following equation (2).

In Equation 2, λ _S represents the model of the speaker S, and X _M represents the speaker's characteristic vector in the case of M subsets. M is the number of subsets selected as having a high recognition rate among the N subsets included in the feature.

The speaker recognition unit 233 recognizes the speaker among voice signal data of all the speakers stored in the voice signal DB 234 based on the maximum likelihood score which is the largest value among the surrounding likelihood scores. In this case, in the voice signal DB, a feature vector extractable from a voice signal of at least one or more speakers and a feature vector of a feature combination including a plurality of features may be calculated and stored for each speaker. For example, the speaker recognition unit 233 may obtain the maximum likelihood score P (X | λ _S ) using Equation 3 below.

은 인식률 가중치를 나타낸다.Where λ _S is the speaker S model, X _M is the speaker's characteristic vector in the case of M subsets,

Denotes a recognition rate weight.

As described above, according to the present invention, a subset of the feature having the highest recognition rate is calculated without calculating a likelihood score for the entire subset included in the characteristic of the speech signal as in the case of applying an advanced missing feature theory (AMFT). By calculating only, the amount of calculation can be reduced and the recognition accuracy can be increased. For example, the recognition error rate in the case of recognizing the speaker using a subset of the recognition rate weights shown in FIG. 4C using N and 8 and 9, and applying the enhanced loss characteristic theory is shown in the following table. You can get it.

Loss characteristic theory AMFT HMFT (N = 8) HMFT (N = 9) Recognition error rate 13.01% 10.34% 10.95%

When the loss characteristic theory of the present invention is called HMFT (Hard-mask Missing Feature theory), the subset N is 8 and 9 in the HMFT method, respectively, 10.34% and 10.95%, respectively, compared to the 13.01% AMFT method. As the recognition error rate is reduced, the recognition accuracy is improved. Unlike the bottom-up method of the AMFT method, which calculates the residual likelihood scores for all subsets from N = 1 to N = 10, the HMFT method calculates the likelihood score from N = 10. Because of the -bottom method, the calculation speed can be much faster.

Meanwhile, an embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by the computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable medium may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes any information delivery media, including computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transport mechanism.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, Therefore, the present invention should be construed as a description of the claims which are intended to cover obvious variations that can be derived from the described embodiments.

100: speaker recognition device
110: voice receiver
120: feature extraction unit
130, 230: speaker matching unit
140: device control unit
231: subset selection unit
232: peripheral likelihood score calculator
233: speaker recognition unit
234: voice signal DB

Claims (12)

A voice receiver for receiving a voice signal;
A feature extractor configured to convert the received voice signal into a frequency domain and extract a feature from the converted voice signal; And
By applying the recognition rate weights to the N subsets included in the extracted characteristic, the M likeness is selected by selecting M subsets higher than a preset recognition rate, and the marginal likelihood score is calculated. A speaker matching unit configured to recognize a speaker who has generated the voice signal based on the result;
The speaker matching unit,
A voice signal DB for storing voice signal data of at least one or more speakers;
A subset selection unit configured to select M subsets having a higher recognition rate than a preset recognition rate by applying a recognition rate weight to the N subsets included in the extracted feature;
A peripheral likelihood score calculator configured to calculate a peripheral likelihood score for the received speech signal using the selected M subsets; And
And a speaker recognizer configured to recognize the speaker among voice signal data of all speakers stored in the voice signal DB based on the maximum likelihood score, which is the largest value among the surrounding likelihood scores. delete The method of claim 1,
And a device controller configured to control an operation of a networked peripheral device in response to the speaker recognition result. The method of claim 1,
The subset selection unit,
The recognition rate weight (

A speaker recognition apparatus using a speech signal for normalizing a value calculated using the following equation, and selecting M subsets having the recognition rate weight higher than a preset recognition rate:

Here, w _N is a weight when the subset is N, R _N is a speaker recognition rate, U _N is a speaker recognition rate. The method of claim 1,
The peripheral likelihood score calculation unit,
Speaker recognition apparatus using a speech signal to obtain the peripheral likelihood score (P (λ _S | X _M )) using the following equation:

Here, lambda _S represents the model of the speaker S, and X _M represents the speaker's characteristic vector in the case of M subsets. The method of claim 1,
The speaker recognizing unit,
Speaker recognition apparatus using a speech signal to obtain the maximum likelihood score (P (X | λ _S )) using the following equation:

Where λ _S is the speaker S model, X _M is the speaker's characteristic vector in the case of M subsets,

Denotes a recognition rate weight. Receiving a voice signal;
Converting the received speech signal into a frequency domain and extracting a characteristic from the converted speech signal; And
By applying the recognition rate weights to the N subsets included in the extracted characteristic, the M likeness is selected by selecting M subsets higher than a preset recognition rate, and the marginal likelihood score is calculated. Matching the speaker that generated the speech signal based on the result;
Matching the speaker,
Selecting M subsets higher than a preset recognition rate by applying a recognition rate weight to the N subsets included in the extracted feature;
Calculating a surrounding likelihood score for the received speech signal using the selected M subsets; And
Recognizing the speaker among voice signal data of all speakers stored in the voice signal DB storing voice signal data of at least one or more speakers based on the maximum likelihood score, which is the largest value among the surrounding likelihood scores. Speaker recognition method using a speech signal. delete The method of claim 7, wherein
And controlling the operation of a networked peripheral device in response to the speaker recognition result. The method of claim 7, wherein
Selecting the subset is,
The recognition rate weight (

2) A speaker recognition method using a speech signal for normalizing a calculated value using the following equation and selecting M subsets having the recognition rate weight higher than a preset recognition rate:

Here, w _N is a weight when the subset is N, R _N is a speaker recognition rate, U _N is a speaker recognition rate. The method of claim 7, wherein
Computing the surrounding likelihood score,
Speaker recognition method using the speech signal to obtain the peripheral likelihood score (P (λ _S | X _M )) using the following equation:

Here, lambda _S represents the model of the speaker S, and X _M represents the speaker's characteristic vector in the case of M subsets. The method of claim 7, wherein
Recognizing the speaker,
Speaker recognition method using a speech signal to obtain the maximum likelihood score (P (X | λ _S )) using the following equation:

Where λ _S is the speaker S model, X _M is the speaker's characteristic vector in the case of M subsets,

Denotes a recognition rate weight.

Images