KR20030075330A - Channel Mis-match Compensation apparatus and method for Robust Speaker Verification system - Google Patents

Abstract

PURPOSE:An apparatus for compensating channel discordant in speaker recognition considering the user mobility and a method therefor are provided to prevent a person recognition rate from decreasing due to channel discordant. CONSTITUTION:An initializing element sets up a center value of each input vector of a voice section of sampled voice data. A center difference calculating element obtains a difference between the mean center value of a common code book and the center value set by the initializing element. A compensating element compensates the difference by using the calculated difference. A speaker registering element registers speakers by using the compensated value.

Description

Channel Mis-match Compensation apparatus and method for Robust Speaker Verification system in speaker recognition considering user mobility

The present invention relates to an apparatus and method for compensating for channel mismatch in speaker recognition in consideration of user mobility. More particularly, the present invention relates to a channel mismatch compensation condition in learning and recognition. By compensating, the present invention relates to a channel mismatch compensation device and method for further improving the success rate of speaker authentication.

The speaker recognition technology may be divided into a speaker verification method and a speaker identification method according to the number of target speakers.

FIG. 1A is a flowchart schematically illustrating an overall process of a conventional speaker identification method, and FIG. 1B is a flowchart schematically illustrating an overall process of a conventional speaker identification method.

First, the speaker confirmation method will be described.

In step S101, when the speaker's voice to be checked is input, in step S103, the feature vector of the input voice is extracted. On the other hand, when the input speaker inputs his ID separately in step S105, in step S106, a speaker model of the input speaker ID is generated.

Subsequently, in step S107, the similarity between the feature vector of the extracted speech and the generated speaker model is calculated. In step S109, it is determined by using a threshold whether the input speech and the speaker ID coincide with each other.

Then, in step S111, the matching result is output, and then ends.

Referring to the speaker identification method shown in Figure 1b as follows.

First, when the voice to be identified is input in step S151, the feature vector of the input voice is extracted in step S153. On the other hand, in step S155, N speaker models are input.

Next, in step S157, the similarity between the feature vector of the extracted input speech and the N speaker models is calculated. In this case, since the similarity calculation is performed for each of the N speaker models with the feature vector of the input speech, the final result is N similarity values.

After selecting the maximum similarity value among the N similarity values in step S159, in step S161, the speaker identification result value is output and ends.

The speaker identification method will be described in more detail as follows.

The speaker verification algorithm follows the process of a general pattern recognition system. That is, when the characteristics representing the characteristics of the object to be recognized are given, it aims to recognize the object based on prior knowledge of the object.

The general pattern recognition system is completed through the steps of training, testing and implementation. In the training phase, the model parameters for the object are estimated while matching the characteristics of the given object with the relationship with the real object. This step is to adjust or adapt the model through data testing. Finally, the implementation step is actually applied, and receives the characteristics of the unknown object to determine the relationship with the modeled object.

Figure 2 is another flow chart schematically showing the overall process of the conventional speaker identification method, which will be described in detail as follows.

The conceptual pattern recognition system is largely divided into a feature extractor and a classifier. A general speaker identification method with a more specific process may be configured as shown in FIG. 2.

(1) Pre-processing process (S201): Pre-processing process is to extract and process a feature parameter to increase the discrimination of the speaker in the system by using the voice information of each speaker coming through the input stage, such pre-processing The process can be subdivided into two steps: extracting and segmenting the input speech segment, and extracting the speech and speaker feature parameters.

In this case, the characteristic parameters include LPC Cepstrum, Delta Chamstrum, Mel Chamstrum, Pitch, Formants, and the like.

(2) Learning and Recognition Process (S203): In the learning and recognition process, the system recognizes and recognizes the speaker-specific feature parameters of the speech extracted in the preprocessing process, and performs a series of processes for confirming the same. This learning and recognition process can be subdivided into a final step of generating a speech and speaker model using a built-up database and applying or deciding a decision rule.

At this time, the learning and recognition process is performed by algorithms such as vector quantization, dynamic time warping (DTW), hidden markov model (HMM), and neural network.

However, the above-described prior arts exhibit proper performance in the speaker's speech pattern by habits or accents that show acquired personality, and have an effect of reducing the false rejection rate at the time of identity verification. There is a problem that decreases the performance of the speaker identification system by increasing the error rate by.

On the other hand, in the case of the speaker identification system, when the certification test is performed with the same system used for learning, the recognition rate is high because there is no difference in characteristics, but when the application is based on the network in real life, many errors occur. Will occur.

That is, when an arbitrary speaker learns an arbitrary system based on a network, generates his own model, and then attempts to authenticate the same, it is not always confirmed with the same system.

Therefore, in this case, the inherent characteristics of a voice input system such as a microphone or a channel such as a sound card are affected, and the bias effect thereof causes a serious misrecognition rate.

On the other hand, there are numerous conventional studies to compensate for such channel mismatches, and these prior arts divide most of the causes of 'inconsistency' into two factors, vocal environmental noise and channel. In general, vocal environmental noise distorts a speech signal in the form of additive noise, adds x (k) to a distorted speech signal, and s (k) to a speech signal before distorting n (k). Speaking of noise, it is expressed as Equation 1 below.

Such additive noise can be effectively removed by the Spectral Subtraction proposed by Steve. In this case, however, the additive noise must be locally stable noise and satisfy the assumption of changing to another stable noise with sufficient time margin.

On the other hand, the distortion of the voice signal due to the channel appears in the form of a convolution with the channel component on the time axis, it is represented by the following equation (2).

Here, h (k) is a channel component, and represents the effect of filtering the voice signal by the channel.

In order to express the channel component convoluted with the voice in this manner, the logarithm of the frequency-converted signal may be taken to represent the spectral region. If the frequency conversion for each signal is X (z), S (z) and H (z), it is expressed by the following formula (3).

In this case, it is assumed that the channel distortion component slowly changes little by little, and if the long-term average of the pure voice cepstrum is 0, the average of the entire interval is calculated and subtracted from the cepstrum region, thereby removing the channel effect. In other words, the effect of the channel appears in the form of an addition to the pure speech cepstruum, so that the estimate of the channel cepstrum can be obtained by averaging the filtered spectrums of the filtered speech, resulting in the estimated channel cepstrum. Can be compensated for the channel effect.

This method is called Cepstrum Mean Subtraction and is used very effectively in speech recognition and speaker recognition for voice over a channel such as a telephone line.

However, the cepstrum mean subtraction method requires acoustic balance of voiced sounds, unvoiced sounds, and ruptured sounds in order for the cepstral mean to be zero for pure speech, so that the speech components other than the channel components are not satisfied. This has the disadvantage of being subtracted.

Therefore, PFCMS (Pole-Filtered Cepstral Mean Subtraction), ACW (Adaptive Component Weighted cepstrum) and Modified Mean Cepstrum Mean Normalization with Frequency Warping (MMCMNFW) methods have been proposed. In addition to the method of directly removing such additive noise or channel components, a method of extracting speech features robust to the effects of noise and channel components has been proposed.

In addition to the existing LPCC and MFCC, PLP (Perceptually Linear Prediction) is known to have robust characteristics that represent the characteristics of speech in a noisy environment by analyzing the auditory characteristics of spectral information. In addition, techniques are known for removing or compensating for noise or channel effects in existing speech features.

Cepstral weighting, especially band-pass filtering, can achieve good performance by relatively attenuating higher-order areas that are sensitive to noise. The previously described CMS is known to be the most effective channel effect normalization method that removes the channel distortion component of the slow-varying cepstrum region in the channel environment. In addition, a technique such as RelAtive SpecTrAL (RASTA), first proposed by Hermansky in 1990, is known to remove channel components by converting the spectrum into a log region and then filtering the spectrum.

The methods discussed so far are processed by the feature extractor, which is the front-end part, before the extracted feature vector is processed by the classifier.

Recently, a method of designing a speaker classifier to be robust to distortion by addition noise and channel component has been proposed and used effectively. Frequently, these methods are classified into model-based channel compensation methods, and are typically represented by using 'Affine Transform'.

In terms of HMM and GMM, which are statistical classifiers commonly used in speech recognition and speaker recognition, changes in noise and channel components are known to shift the characteristics of speech expressed in mean and variance in vector space. This movement can be expressed by scaling, rotation, and translation, and is generally used in a stereo voice database that simultaneously collects voices distorted by voice, noise, and channel components before being distorted. This conversion process is estimated. The estimated conversion process can be applied to the classifier to model the input speech to be robust to noise and channel components.

However, this method also has a problem in that it is necessary to construct a stereo voice database capable of estimating various conversion processes.

An object of the present invention for solving the problems of the prior art as described above to improve the success rate of the speaker authentication by compensating for the difference between the average value of the common codebook center and the average value of the individual codebook center in the channel mismatch conditions during learning and recognition It is to provide an apparatus and method for compensating for channel mismatch.

Figure 1a is a flow chart schematically showing the overall process of the conventional speaker identification method,

Figure 1b is a flow chart schematically showing the overall process of the conventional speaker identification method,

Figure 2 is another flow chart schematically showing the overall process of the conventional speaker identification method,

FIG. 3 is a diagram illustrating a concept of compensating a difference in advance between a center value of a common codebook and a center value of learning data in advance according to an embodiment of the present invention;

4 is a flowchart illustrating a channel mismatch compensation method in speaker speech learning according to an embodiment of the present invention.

5 is a flowchart illustrating a method of compensating for channel mismatch during speaker voice confirmation according to an embodiment of the present invention.

In order to achieve the above object, according to the present invention, in the apparatus for compensating for channel mismatch in learning in a speaker learning system in consideration of user mobility, each center value of all input vectors of speech sections of the input sampled speech data is determined. Initialization means for setting; The difference between the average center value of the common codebook and the center value set by the initialization means ( Means for calculating a center value difference; Compensation means for compensating for the difference between the mean center value of the common codebook and the center value for the input vector using the center value difference calculated by the center value difference calculating means; And speaker registration means for performing a speaker registration process using a value compensated by the compensation means. The apparatus for compensating for channel inconsistency in learning in a speaker identification system considering user mobility is provided.

A method of compensating for channel mismatch in learning in a speaker learning system in consideration of user mobility, the method comprising: a first step of setting respective center values of all input vectors of speech sections of input sampled speech data; The difference between the mean center value of the common codebook and the center value set in the first step ( Obtaining a second step; A third step of compensating the difference between the mean center value of the common codebook and the center value of the input vector using the center value difference calculated in the second step; And a fourth step of performing a speaker registration process using the value compensated in the third step. The channel mismatch compensation method of learning in a speaker learning system considering user mobility is provided.

In addition, in the apparatus for compensating for channel inconsistency in a speaker identification system considering user mobility, an average value of an input vector of a silent section of input sampled speech data is obtained, and a weight is determined by a sigmoid function. Weight determination means; Center value setting means for setting one center value for all the input vectors of the voice interval of the input sampled speech data; Center value difference calculating means for obtaining a difference between an average center value of a common codebook and a center value set by the center value setting means; Compensation means for compensating the input vector using the difference in the center value calculated by the center value difference calculating means and the weight determined by the weight determining means; And a speaker identification means for performing a speaker identification process by using the input vector compensated by the compensation means.

Also, in a method of compensating for channel mismatch in confirmation in a speaker identification system considering user mobility, an average value of an input vector of a silent section of input sampled speech data is obtained, and a weight is determined by a sigmoid function. A first step of making; A second step of setting respective center values for all input vectors of voice sections of the input sampled voice data; Obtaining a difference between the average center value of the common codebook and the center value set in the second step; A fourth step of compensating the input vector using the difference between the center value calculated in the third step and the weight determined in the first step; And a fifth step of performing a speaker identification process by using the input vector compensated in the fourth step.

Hereinafter, an apparatus and method for compensating for channel mismatch in speaker recognition considering user mobility according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The proposed method removes the channel mismatch condition between learning and recognition as a compensation for the difference between the mean value of the common codebook center and the mean value of the personal codebook center.

FIG. 3 is a diagram illustrating a concept of compensating a difference in advance between a center value of a common codebook and a center value of learning data according to an embodiment of the present invention.

The characteristic parameter of the voice spoken by the speaker varies greatly depending on the characteristics of the connecting system, that is, the sound card, the microphone, and the background noise. This characteristic parameter change causes a difference from the speaker area, resulting in a self rejection rate, as shown in FIG.

However, this difference does not significantly affect the distribution form of the speaker area. That is, the change of the characteristic parameter distorted by the channel affects the mean value of the speaker region, but does not affect the variance value. Therefore, the technique proposed in the present invention compensates the difference between the center value of the common codebook and the center value of the learning data in advance by learning, as shown in FIG. By compensating and confirming the difference by order with the value, it resolves the sudden drop in the recognition rate caused by the channel mismatch.

However, since the unconditional average value compensation results in the authentication error of the impersonator, the average value compensation through appropriate weights proportional to the channel variation is necessary. Therefore, in consideration of the distribution of the silent section except for the speech section, the difference between the channel and the channel during the training is compensated by the weight of the nonlinear function.

4 is a flowchart illustrating a channel mismatch compensation method in speaker speech learning according to an embodiment of the present invention.

(1) Step S401: Initialization process. One center value is set for all the input vectors of the voice interval, and this is expressed by the following equation (Equation 4).

here, Is the number of all input vectors.

(2) Step S402: A difference (Bias) between the average center value of the common codebook and the center value of the input vector is obtained. The following expression is expressed by Equation 5 below.

Where p = 0, 1, ..., k.

(3) Step S403: The difference between the mean center value of the common codebook and the center value with respect to the input vector is compensated for. The following expression is expressed by Equation 6 below.

(4) Step S404: A speaker registration process is performed using the difference of the compensated center value.

5 is a flowchart illustrating a method of compensating for channel mismatch during speaker voice confirmation according to an embodiment of the present invention.

(1) Step S501: Initialization process. The average value of the input vector of the silent section is obtained, and the weight is determined by a sigmoid function, which is represented by Equation 7 below.

here, Is background noise at the time of registration of personal speaker, Is background noise, and w is a weight to be obtained in this step S501.

(2) Step S502: One center value for all the input vectors of the speech section is set, and this is expressed by the following formula (8).

here, Is the input vector, Is the number of all input vectors.

(3) Step S503: The difference between the average center value of the common codebook and the center value of the input vector is obtained. The following expression is expressed by Equation 9 below.

Where p = 0, 1, ..., k.

(4) Step S504: The difference between the mean center value of the common codebook and the center value of the input vector is compensated by giving the above weight, which is expressed by the following formula (10).

(5) Step S505: A speaker verification process is performed.

On the other hand, if the learned microphone is the same as the microphone used for the certification test, the channel recognition is not applied or the channel compensation algorithm is applied. Showed a significant performance difference between.

The results of these experiments are shown in Table 1 below.

TABLE 1

As shown in Table 1, when the authentication experiment with the microphone 4 for the model learned with the microphone 1, it can be seen that the identity authentication rate is greatly improved from 20% to 90%.

While the invention has been described above based on the preferred embodiments thereof, these embodiments are intended to illustrate rather than limit the invention. It will be apparent to those skilled in the art that various changes, modifications, or adjustments to the above embodiments can be made without departing from the spirit of the invention. Therefore, the protection scope of the present invention will be limited by the appended claims, and should be construed as including all such changes, modifications or adjustments.

As described above, according to the present invention, in the speaker authentication method, the difference between the center value of the common codebook at the time of learning and the center value of the learning data is compensated for in advance, and the center value of the common codebook and the center of the learning data are also verified. By compensating and confirming the difference by order with the value, there is an effect of resolving a sudden drop in the recognition rate caused by the channel mismatch.

Claims (16)

A device for compensating for channel mismatch in learning in a speaker learning system considering user mobility, Initialization means for setting respective center values for all input vectors of speech sections of the inputted sampled speech data; The difference between the average center value of the common codebook and the center value set by the initialization means ( Means for calculating a center value difference; Compensation means for compensating for the difference between the mean center value of the common codebook and the center value for the input vector using the center value difference calculated by the center value difference calculating means; And And speaker registration means for performing a speaker registration process by using the value compensated by the compensation means. The method of claim 1, The initialization means, The apparatus for compensating for channel mismatch in learning in a speaker learning system considering user mobility, wherein the center value is determined by Equation 1 below. [Equation 1] here, Is the input vector, Is the number of all input vectors. The method of claim 1, The compensation means, Compensation apparatus for a channel mismatch in learning in a speaker learning system considering user mobility, wherein the difference of the center value is compensated by Equation 2 below. [Equation 2] here, Is the compensated value, Is the difference between the mean center of the common codebook and the center of the input vector. In the channel mismatch compensation method for learning in the speaker learning system considering the user mobility, A first step of setting respective center values for all input vectors of speech sections of the input sampled speech data; The difference between the mean center value of the common codebook and the center value set in the first step ( Obtaining a second step; A third step of compensating the difference between the mean center value of the common codebook and the center value of the input vector using the center value difference calculated in the second step; And And a fourth step of performing a speaker registration process using the value compensated in the third step. The method of claim 4, wherein The first step is, The channel mismatch compensation method for learning in the speaker learning system considering the user mobility, characterized in that the center value is determined by the following [Equation 3]. [Equation 3] here, Is the input vector, Is the number of all input vectors. The method of claim 4, wherein The third step, Compensation for the channel mismatch during learning in the speaker learning system considering the user mobility, characterized in that the difference of the center value is compensated by the following [Equation 4]. [Equation 4] here, Is the compensated value, Is the difference between the mean center of the common codebook and the center of the input vector. A device for compensating for channel mismatch in confirmation in a speaker identification system considering user mobility, Weight determination means for obtaining an average value of the input vector of the silent section of the input sampled speech data and determining the weight by a sigmoid function; Center value setting means for setting one center value for all the input vectors of the voice interval of the input sampled speech data; Center value difference calculating means for obtaining a difference between an average center value of a common codebook and a center value set by the center value setting means; Compensation means for compensating the input vector using the difference in the center value calculated by the center value difference calculating means and the weight determined by the weight determining means; and And a speaker identification means for performing a speaker identification process using the input vector compensated by the compensation means. The method of claim 7, wherein The weight determining means, The apparatus for compensating for channel inconsistency in confirmation in a speaker identification system considering user mobility, wherein the weight is determined by Equation 5 below. [Equation 5] here, Is background noise at the time of registration of personal speaker, N_bg ^ (p) is Background Noise and w is a weight to be obtained. The method of claim 7, wherein The center value setting means, The apparatus of claim 2, wherein the center value is determined by Equation 6 below. [Equation 6] here, Is the input vector, Is the number of all input vectors. The method of claim 7, wherein The compensation means, Compensation apparatus for a channel mismatch in confirmation in a speaker confirmation system considering user mobility, which is compensated by Equation 7 below. [Equation 7] here, Is the compensated value, Is the result value of the said center value difference calculating means. In the channel mismatch compensation method in the speaker identification system considering the user mobility, A first step of obtaining an average value of an input vector of a silent section of input sampled speech data and determining a weight by a sigmoid function; A second step of setting respective center values for all input vectors of voice sections of the input sampled voice data; Obtaining a difference between the average center value of the common codebook and the center value set in the second step; A fourth step of compensating the input vector using the difference between the center value calculated in the third step and the weight determined in the first step; And And a fifth step of performing a speaker identification process by using the input vector compensated in the fourth step. The method of claim 11, The first step is, The method of claim 2, wherein the weight is determined by Equation 8 below. [Equation 8] here, Is background noise at the time of registration of personal speaker, Is the background noise and w is the weight to obtain. The method of claim 11, The second step, The channel mismatch compensation method when confirming in the speaker confirmation system considering the user mobility, characterized in that the center value is determined by the following [Equation 9]. [Equation 9] here, Is the input vector, Is the number of all input vectors. The method of claim 11, The fourth step, A channel mismatch compensation method when confirming in a speaker identification system considering user mobility, which is compensated by Equation 10 below. [Equation 10] here, Is the compensated value, Is the result value of the said center value difference calculating means. A computer-readable recording medium capable of executing a channel mismatch compensation program when learning in a speaker learning system considering user mobility, A first step of setting respective center values for all input vectors of speech sections of the input sampled speech data; The difference between the mean center value of the common codebook and the center value set in the first step ( Obtaining a second step; A third step of compensating the difference between the mean center value of the common codebook and the center value of the input vector using the center value difference calculated in the second step; And And a fourth step of performing a speaker registration process using the value compensated in the third step. The center value is determined by Equation 11 below, and the difference in the center value is compensated by Equation 12 below. A computer-readable recording medium capable of running a computer. [Equation 11] here, Is the input vector, Is the number of all input vectors. [Equation 12] here, Is the compensated value, Is the difference between the mean center of the common codebook and the center of the input vector. A computer-readable recording medium capable of executing a channel mismatch compensation program upon confirmation in a speaker identification system considering user mobility, A first step of obtaining an average value of an input vector of a silent section of input sampled speech data and determining a weight by a sigmoid function; A second step of setting respective center values for all input vectors of voice sections of the input sampled voice data; Obtaining a difference between the mean center value of the common codebook and the center value set in the second step; A fourth step of compensating the input vector using the difference between the center value calculated in the third step and the weight determined in the first step; And And a fifth step of performing a speaker identification process using the input vector compensated in the fourth step. The weight is determined by Equation 13 below, the center value is determined by Equation 14 below, and compensated by Equation 10 below. A computer-readable recording medium capable of executing a channel mismatch compensation program upon verification at. [Equation 13] here, Is background noise at the time of registration of personal speaker, Is the background noise and w is the weight to obtain. [Equation 14] here, Is the input vector, Is the number of all input vectors. [Equation 15] here, Is the compensated value, Is the result value of the said center value difference calculating means.