KR20010036358A - Text-Independent Speaker Identification Using Telephone Speech - Google Patents

Abstract

PURPOSE: A sentence independent type speaker discriminating method using a telephone voice is provided to increase a feature of a speaker and restrain a channel distortion. CONSTITUTION: A matrix is obtained from a learning data of speakers of several environment according to an independent component analysis(ICA) estimation algorithm. An average is subtracted from an input cepstrum. The subtracted result is converted into the obtain matrix. A new conversion cepstrum is used in a study and discrimination.

Description

Text-Independent Speaker Identification Using Telephone Speech}

The present invention relates to a speaker identification method of telephone speech. In particular, a method of transforming feature parameters into a space robust to channel variations of a phone voice is directed to a speaker identification method using a cepstrum vector transformation using independent component analysis.

Advances in speech information processing technology and improved computing power of computers have enabled the construction of probabilistic models that can accommodate large-scale speech data processing and various speech variations. Accordingly, the speaker recognition technology has also made great strides.

Telephone speech, which is the main field of speaker recognition, cannot accurately estimate and compensate because the unpredictable nonlinear distortion is reflected in the voice due to the limitation of bandwidth and the channel response characteristic of the transmission line, which causes severe performance degradation. Therefore, for the practical use of speaker recognition, a method of estimating channel distortion occurring in telephone voice and compensating for it is essential.

1 shows a distortion process of telephone voice.

Voices spoken by one speaker are collected by the microphone and transmitted to the other speaker by telephone line. In this process, an effect of filtering the speech signal occurs, which is represented by a spectral tilt in the frequency axis.

Therefore, when learning and recognizing with voices collected in different environments, differences in feature parameters appear non-additively, leading to severe degradation of recognition rate.

The spectrum of the telephone voice is represented by the transmission function of the transmission line by the spectrum of the voice as shown in Equation 1 through the process as shown in FIG.

From here Are transfer functions that represent the basic vibration of the vocal cords, the characteristics of the saints, the characteristics of the collecting microphone, and the characteristics of the transmission line.

In the cepstrum domain after the conversion of Equation 2, the product of these filter functions is represented as a linear sum.

The glottal pulse of the vocal cords during the analysis Is excluded. Therefore, the characteristic of voice is that Since it is expressed only by, it is necessary to suppress the characteristic functions of the transmission environment in order to reduce the distortion between the learning speech and the recognition speech.

As a method of compensating for such channel distortion, assuming that the channel characteristic does not change in a call, methods for estimating the distortion as a fixed value in a long period of speech have been disclosed.

Representative examples disclosed above include Cepstral Mean Subtraction (CMS), Signal Bias Removal (SBR), and Affine Transform of Cepstrum (ATC).

In the cepstrum mean subtraction method (CMS), when the types of phonemes included in the speakers' uttered sentences are distributed evenly, the cepstrum parameters that are the result of the analysis are evenly distributed in the sentences. Therefore, it can be assumed that the average values of the cepstrums of the voices are the same.

If it is distorted by the transmission function of the transmission line, only the average value is affected since the characteristics of the channel are hardly changed in one call. Therefore, it is assumed that the average of the cepstral vectors of the non-distorted speech is 0, and the average of the cepstral vectors of the long interval speech is estimated as the channel distortion.

The subtracted cepstrum vector is used for learning and recognition. In other words, this method ignores the static distribution of parameters in vector space and attempts to reflect only the difference from the mean that is not distorted by the transmission function in the feature parameter.

2 shows a simple process of the cepstrum mean subtraction method. Both learning and cognition use cepstrum that has undergone the same deduction process, and there is no need for special background information estimation process when channel distortion by transmission line is constant.

As described above, the cepstrum mean subtraction method subtracts only the average of long-range speech, so that the calculation amount is not large and simple as shown in the following equation, and it shows stable results in general telephone speech processing.

In addition, it is possible to obtain better results in case of long sentence free speech recognition.

SBR (SBR) assumes that a constant deviation occurs between the learning data and the recognition voice due to channel distortion, like the CMS, and uses the maximum likelihood estimation (MLE) for the speech data used for learning the recognizer. The constant channel deviation is estimated and subtracted from the input vector so that the distribution of and feature vectors is most similar. In fact, it has been applied to telephone speech recognizer using discrete HMM, which shows high performance.

Assuming that channel distortion is fixed in the spectral region, a constant deviation vector that minimizes channel distortion for the codebook of the entire training material. The value of is as follows.

z (·): center vector of input by codebook of study material

Repeated Channel Distortion by Maximum Likelihood Estimation Estimate of Find it.

3 shows a process using a signal deviation cancellation method. In the signal deviation cancellation method, unlike in the spectral mean subtraction method, the parameters used for learning and recognition go through different processes.

The learning is divided into two processes: the speaker model learning and the cepstrum codebook learning. The speaker model is trained using non-normalized cepstrum, and the codebook of the learning environment is obtained by k-means algorithm with the same training data. In the recognition process, the optimal deviation from the learning environment codebook that minimizes the distortion of the input chord for the learning data is obtained, and the chordstrum that subtracts the deviation is used for recognition. That is, the output probability of the input voice is an optimal value that allows horizontal and vertical variations in space.

This method estimates the best channel function for the learning materials, so it is less affected by the length of the input voice. In addition, if the distortion of the learning material is less accurate, it is possible to estimate more accurately, it is possible to reduce the loss of information by the normalization than the cepstrum mean subtraction method.

In the conventional channel distortion compensation method, the cepstrum mean subtraction method (CMS) has a low reliability of the estimated value when the length of the input voice is short or the distribution of the phonemes is uneven in the voice, and when the influence of the channel in the voice is not large. The speaker's own static information is lost, causing performance degradation.

In addition, when the distortion by the transmission line is not constant and changes with time, estimation and subtraction of the distortion are impossible.

In addition, since the SBR follows the assumption that the channel characteristics are fixed for all speech sections, it is impossible to estimate the distortion that changes with time.

In addition, when the distortion of the input voice is large, there is a problem that it is difficult to accurately estimate the channel distortion.

The compensation methods assume that the channel distortion function does not change within the entire call, and applies the same channel distortion compensation parameter to the entire call. However, in addition to the dynamic characteristics such as personal utterance, dialect, and personal language, the speaker's characteristics such as vocal vocal cords, vocal cords, and nasal characteristics appear constant for the entire utterance as well as the channel characteristics.

Therefore, in the process of removing the channel characteristic, the loss of the speaker's unique static characteristics occurs, and there is a problem in that it is impossible to remove the channel distortion which is not static and exhibits variation over time.

The present invention is to solve the problems of the prior art, an object of the present invention is to increase the speaker characteristics common in speech as well as compensation of channel distortion, robust robust speaker extraction method using independent component analysis to suppress channel distortion To provide.

As a technical idea for achieving the object of the present invention, as a method of mapping a feature vector of speech to a new feature space using independent component analysis, several feature functions of the cepstrum vectors of the speech signals collected in various environments Is assumed to be a linear combination of

A method of using separated cepstrum vectors for learning and recognition using independence analysis is presented.

In the new vector region of the present invention, feature information of repeated voices is emphasized, and random channel distortions are suppressed to improve discrimination between speakers in the cepstrum space, and to appear as a more robust space against channel environment changes. have.

1 is an explanatory diagram illustrating a distortion process of telephone voice.

2 is an explanatory diagram for a Cepstrom average subtraction method (CMS).

3 is an explanatory diagram for a signal deviation cancellation method (SBR).

4 is an explanatory diagram illustrating a mixing process of an input cepstrum, which is a main process of the present invention.

5 is an explanatory diagram of a domain transformation effect by independent component analysis, which is a main process of the present invention.

6 is an explanatory diagram for a Cepstrum transformation process by independent component analysis of the present invention.

7 is a speaker distribution diagram of the original cepstrum space before conversion.

8 is a speaker distribution diagram of a cepstrum space by a linear transformation.

9 is an explanatory diagram for a pretreatment process by independent component analysis.

Hereinafter, the configuration and operation of the embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The present invention proposes a robust speaker recognition method in a severely distorted environment using independent component analysis, which has been studied in the field of signal processing in recent years, and describes the implementation of the system.

Independent component analysis (ICA) refers to a method of effectively separating a signal having a different characteristic when it is linearly mixed by using a statistical method. Applied in many fields, the results are good.

The term independent component analysis was first proposed by Herault and Jutten in 1986 and was named after similarity to the existing principal component analysis (PCA).

Independent component analysis distinguishes not only the correlation but also the dependency between signals. Therefore, independent component analysis can be seen as a generalized form of principal component analysis.

Independent component analysis can define statistical dependencies using higher-order statistics (HOS), information theory, etc., and its independence by neural network learning methods that estimate linear weights. You can estimate and separate people. In this case, depending on the method of defining the dependency, the characteristics of the independent component analysis appear in slightly different forms.

Typical applications of independent component analysis include voice processing such as noise removal, separation, and sound quality improvement, various image processing fields, and medical fields analyzing computed tomography images.

In addition, it has been applied to general pattern recognition, such as extracting representative feature values from complex patterns or emphasizing ambiguity.

In the present invention, as a method of mapping a feature vector of speech to a new feature space using independent component analysis, a method of improving the performance of a speaker identifier of a severely distorted telephone speech is proposed.

First, the cepstrum vector space representing the speaker's characteristics is transformed into a space robust to the reduction of the discriminant power due to channel distortion. In the newly transformed space, the features of speech are emphasized, and distributions with indistinguishable distinction between speakers can be distinguished from each other. Therefore, in the case of a severely distorted voice such as a telephone voice, it is possible to obtain a new feature space having a strong characteristic in reducing the discrimination of the speaker space due to lost information.

In other words, an embodiment of the present invention is a method for transforming feature parameters into a space robust to channel variation of a telephone voice, and employs cepstrum vector transformation using independent component analysis (CVTICA). Implemented by the method.

In more detail, ICA (independent component analysis) is described. A random vector composed of linearly mixed two or more signals having different characteristics into independent signals according to a statistical method. It is also known as separating. In other words, it is assumed that arbitrary signals are mixed multiplied by the weights of several signals, and the dependencies between the mixed signals are defined based on information theory.

In addition, according to a learning method used for neural networks, statistically independent signals may be obtained by estimating a weight at which the dependency between signals is minimal. In particular, it is drawing attention as a solution to the cocktail party problem that separates the mixed voice signals into individual voices.

first Random variables Enter Probability-independent variables It is assumed that the linear combination of these is made. q are random variables that are not less than p and whose mean is zero. These random variables are two random vectors Wow Constitute Is And mixed matrix It is expressed as the product of. It is assumed that some errors and constant deviations are added.

The purpose of the independent component analysis is to And linear mixed matrix Is to obtain an estimate of. To Estimate of, To Equation 6 can be replaced by the following equation.

therefore, Instead, the observation signal is converted into an independent signal. The solution of the independent component analysis can be obtained by

Statistical independence can be calculated as the difference between the estimated entropy of combined entropy and each entropy, which is defined as mutual information.

In addition, the mutual information may be calculated using negative normalized entropy estimated by a normal distribution having a covariance and an average such as the estimated components.

Is Is the covariance matrix of Is the entropy according to Shannon's definition. The weights that minimize the mutual information defined in this way can minimize the probability of combining each component, and the separation matrix at this time Are coefficients that maximize independence between components. In addition, negtropy can be estimated by a nonlinear higher-order cumulative distribution function. These estimation functions Expressed as

Is defined as the contrast function. The separation matrix of the independent components can be obtained from the estimated Neg entropy using the contrast function of Equation 9, and the separated independent components are characterized by the contrast function.

The following describes the linear mixing process of the cepstrum of the present invention.

In the method of the present invention, the independent component analysis, which is mainly used to separate the mixed signals, is applied to the spectral parameter, which is the speaker's feature vector, to extract a feature robust to channel distortion. In other words, by using independent component analysis that emphasizes a common signal but suppresses a different signal, a linear transformation is estimated to emphasize a voice signal and suppress a randomly appearing channel signal in a general telephone voice.

In addition, when a new spatial cepstrum vector is used for learning and identification, distinction of speakers can be made clear and distortion of the speaker model by channels can be minimized.

4 shows the mixing process of the input cepstrum.

FIG. 5 shows the effect of region transformation by independent component analysis, (a) shows wrong speaker modeling due to channel distortion and overlapping space, and (b) shows speaker modeling clearly distinguished by separation of each component. .

To recapitulate the assumptions of the conventional channel compensation methods described above,

The first is that the nature of the distorted function of the channel changes depending on the connection path that changes with each call. However, because the speaker identifier cannot obtain accurate information about the channel, it is practically impossible to calculate the correct channel distortion functions.

Secondly, the characteristics are almost constant within a connected call. In other words, channel distortion is constant throughout the call and changes very slowly compared to the change in voice. Again, it is not known how many different channel characteristic functions are included in the telephone voice, and the distinction between the transfer functions representing the existing voice characteristics is not clear.

Therefore, it is assumed here that the number is any fixed number. That is, Equation 2 representing the cepstrum of telephone speech may be expressed as a sum of characteristic functions of arbitrary numbers as follows.

In accordance with Equation 10, the cepstrum is expressed as a sum of various speech transfer and channel distortion functions. And the resulting chop strum It consists of four different components. Thus, these statistically distinct functions are defined in vector space Assume a linear combination of two functions.

Assumption 1.

The p-order cepstrum is a linear mixture of p statistically independent transfer functions representing the character of the saints.

In practice, there may be more or less than this, and the distinction is not clear. By hypothesis 1, the cepstrum can be thought of as one of the linear mixed models of equation (6).

To identify, the vocal tract transfer function that characterizes the voice in Only important and channel distortion Should be suppressed. The channel transfer functions included in the cepstrum are fixed within one call but are not related between different calls. It also has the characteristic of changing slowly little by little within a currency.

Therefore, the average of the channel transfer functions included in the cepstrum of telephone speech collected in a sufficient number of environments can be regarded as Gaussian random noise following a Gaussian distribution by normalizing to zero for one call.

Home 2

The dynamic variation of the channel distortions contained in the phonetic spectral vectors has the property of noise following the Gaussian distribution.

That is, Is average Normalized Channel Distortion Function This is assumed to be Gaussian noise. However Transfer function of voice included in Has common features in different currencies.

Therefore, when the cepstrum vectors of the phonetic voices of various environments undergo independent component analysis, the common voice feature functions are clearly separated in space, and the channel functions are regarded as added noise and suppressed.

Therefore, a new cepstrum space that is robust to environmental variations as shown in FIG. 5 can be obtained. In addition, the independent component analysis estimates the linear transformation matrix transformed around the origin, and thus, it is difficult to obtain an optimal linear separation matrix when constant channel distortion exists in the input cepstrum space. Accordingly, the linear transformation matrix is estimated after subtracting the average so that the overall mean of the cepstrum space becomes zero.

The following describes the conversion process, which is the main process of the embodiment of the present invention.

Figure 6 shows the Cepstrum transformation process (CVTICA) by the independent component analysis of the present invention.

In order to process t of Gaussian 2 as Gaussian noise, a mixed matrix must be obtained from speech of general distortion. And we perform the learning and identification with the strong spatial cepstrum obtained through the inverse transformation of the transformation. This robust space cepstrum is Assume In fact, the best mixed matrix Since we cannot find, its inverse transformation matrix through independent component analysis Estimate The estimated matrix is then the inverse of the mixed matrix. Rather, matrices of different ordering components and different sizes are estimated.

In the above formula Is an exchange matrix that changes the order of the rows, Is a diagonal matrix that fits the size of the rows. In the independent component analysis, the channel characteristics are considered as Gaussian noise t [n] with a mean of 0 and suppressed. Will be obtained.

Therefore, the linear transform cepstrum obtained by independent component separation matrix in general telephone speech of hypothesis 1 Can be obtained by the robustness of channel variation and high discrimination of vector space. Then, in order to use the new cepstrum as a feature vector in a probabilistic model such as HMM, it is necessary to define the distance between the two vectors in the new region. The relationship between the distance in the new cepstral vector region linearly transformed by the cepstral distance before the transformation can be obtained from the following equation.

That is, the matrix when computing the distance between two We can get a distance measure similar to the Mahalanobis distance, which is multiplied by two vectors. Mahalanobis street In the form of Is a covariance matrix, usually approximated diagonally. Thus, the distance scale in Equation 13 above is a special form of Mahalanobis distance.

The above processes can be summarized as follows algorithm CVTICA.

a. Transformation Matrix According to Independent Component Analysis Estimation Algorithm in Speakers' Learning Materials of Various Environments Obtain

b. ← : Subtract the mean from the input column.

c. ← Convert to the matrix found in a.

d. New transformation cepstrum in learning and identification Use

7 and 8 show the original cepstrum The cepstrum converted by That is, FIG. 7 shows the speaker distribution diagram of the original cepstrum space, and FIG. 8 shows the speaker distribution diagram of the cepstrum space by linear transformation.

In FIG. 7 and FIG. 8, '+' indicates a cepstrum of a male speaker, and 'o' indicates a cepstrum of a female speaker. Only the fifth and sixth component values are expressed in two dimensions.

In the case of Figure 7, the entire distribution is not gathered, but spread a lot, the distinction between the two speakers is not clear. On the other hand, in the case of Figure 8, the distribution of each speaker is shown a lot of form and the discrimination between the two speakers also shows a greatly improved form. Therefore, it can be seen that the method of the present invention is effective for improving speaker discrimination.

CMS, SBR, and linear transformation estimation methods described above are new These are special forms of and are distinguished by transformation matrices and constant deviation estimation methods.

Existing channel normalization methods suppress the channel signal and emphasize the voice signal using the characteristics of the voice signal. The method of the embodiment of the present invention increases the discrimination ability of the vector space and suppresses the channel signal from a statistical and pattern analysis point of view.

9 shows a preprocessing part of the built-in speaker identification system of the present invention. Transformation matrix to independent columns Uses the entire speaker's learning data to estimate the independent component matrices to absorb variations for any input cepstrum. The transformed independent cepstrum is used in the same way for both learning and perception.

For modeling speaker distribution, we use the HMVQM (hidden Markov VQ-codebook model), a special form of continuous HMM.

Unlike continuous HMMs, which estimate the observed probabilities of HMMs as the weighted sum of the normal distributions, this model defines the minimum vector quantization distance as

To construct a statement-independent speaker identification system, the HMM's connection structure uses an ergodic topology where connections exist between all state pairs. The possible state transitions in the learned HMM can stochasticly map all phoneme sequences occurring in the learning material. As a result, state transitions map major phonemes.

As described above, the conventional channel compensation methods assume that channel distortion is fixed for one pass, and thus may cause loss of static speaker information in the entire speech, and it is impossible to solve channel distortion that changes over time. In view of this, we can learn how to improve the performance of the speaker identifier in speech that is severely distorted, such as telephone voice, by learning and recognizing the cepstrum vector, which is the speaker's characteristic space, with the transformed spectrum through independent component analysis. There is an effect of significantly improving the speaker's recognition rate.

In order to prove the effect, the experiment was conducted with SPIDRE, which is a real long distance voice data widely used in the sentence-independent speaker recognition study.

Also, in order to verify the effect of channel compensation, we classified the learning and experimental data into the same and different cases.

Conventional methods have the recognition rate up to 10% higher than that of the basic system without channel compensation in the same channel, whereas the proposed method has not deteriorated its performance. This was also seen when it was improved.

In addition, in the case of a different channel, both the existing channel compensation methods and the proposed method have significantly improved the recognition rate compared to the basic system that does not perform channel compensation. In addition, the proposed method improves the recognition rate from 9% to 15% in all system environments compared to the existing channel compensation methods.

This is because the method of the present invention transforms into a more discreet space while generalizing the existing methods, thereby preventing erroneous speaker modeling due to instantaneous channel distortion, and thus, has better robustness against channel environment changes.

Since the speaker identification method of the present invention deals with the conversion in the general feature space, it can be applied not only to the speaker recognizer but also to other speech recognition fields.

Claims (10)

In the method of identifying a speaker by using a voice signal, Assuming that the spectral vectors of the speech signals collected in various environments are linear combinations of a predetermined number of feature functions; Linear transforming the feature functions into a new feature space using independent component analysis; And a process of using the new spectral vectors of the new space for learning and recognition. The method according to claim 1, In the linear transformation process, the speech signal repeatedly appearing using the independent component analysis is emphasized, and the channel distortion signal appearing randomly is suppressed to improve the discrimination between speakers in the cepstrum space and to make the space more robust to the channel environment change. Sentence independent speaker identification method using a telephone voice, characterized in that the implementation process. The method according to claim 1, The cepstrum vector is separated using independent component analysis, the common signal is emphasized, and the different signals are suppressed sentence independent speaker, characterized in that the suppressed. The method according to claim 1, And said feature functions comprise a plurality of channel distortion functions and a plurality of voice feature transfer functions. The method according to claim 1, The linear combination of the feature functions is a sentence independent speaker identification method using a phonetic voice, characterized in that defined by the following equation expressed as the sum of the characteristic function of any number. The method according to claim 4, h [n] is a speech transfer function, and t [n] is a channel distortion function. The method according to claim 1, The linear transformation process is a sentence-independent speaker identification method using telephone speech comprising applying an independent component analysis to a spectral parameter, which is a speaker's feature vector, to extract features robust to channel distortion. The method according to claim 1, The linear transformation process is a sentence-independent speaker identification method using a phonetic voice comprising the use of independent component analysis to obtain the speaker modeling clearly separated from each component in the speaker modeling due to the overlap of the channel distortion and space. A method of identifying a speaker by assuming that the spectral vectors of voice signals collected in various environments are linear combinations of a predetermined number of feature functions, (a) the process of obtaining the transformation matrix from the speaker's learning data according to the independent component analysis estimation algorithm; (b) subtracting the mean from the input cepstrum, (c) converting the matrix obtained in step (a) to (d) A sentence-independent speaker identification method using telephone speech, including the use of new transformed cepstrums for learning and identification. The method according to claim 9, The conversion matrix is a character-independent speaker identification method using a telephone voice, characterized in that defined by the following equation.