CN107301859B - Speech conversion method under non-parallel text condition based on adaptive Gaussian clustering - Google Patents

Abstract

The invention discloses a voice conversion method under the condition of non-parallel text based on adaptive Gaussian clustering, and belongs to the technical field of voice signal processing. Firstly, the method based on the combination of unit selection and channel length normalization is used to align the speech feature parameters of the non-parallel corpus, and then the adaptive Gaussian mixture model and the bilinear frequency bending plus amplitude adjustment are trained, and the speech conversion method is obtained. The required conversion function is finally used to achieve high-quality speech conversion. The invention not only overcomes the limitation of requiring parallel corpus in the training stage, realizes the speech conversion under the condition of non-parallel text, and has stronger applicability and versatility, but also uses the adaptive Gaussian mixture model to replace the traditional Gaussian mixture model, and solves the problem of the Gaussian mixture model. The problem of inaccuracy in the classification of speech feature parameters, and the combination of adaptive Gaussian mixture model and bilinear frequency bending and amplitude adjustment is better in the transformation of personality similarity and speech quality.

Description

Speech conversion method under non-parallel text condition based on adaptive Gaussian clustering

technical field

The invention relates to a voice conversion technology, in particular to a voice conversion method under the condition of non-parallel text, and belongs to the technical field of voice signal processing.

Background technique

Speech conversion is an emerging research branch in the field of speech signal processing in recent years. It is carried out and developed on the basis of speech analysis, recognition and synthesis research.

The goal of speech conversion is to change the voice personality of the source speaker so that it has the voice personality of the target speaker, that is, to make the voice spoken by one person sound like the voice spoken by another person after conversion, while preserving the semantics. .

Most speech conversion methods, especially GMM-based speech conversion methods, require that the corpus used for training be parallel text, that is, the source speaker and the target speaker need to emit sentences with the same speech content, speech duration, and pronunciation rhythm and rhythm. Emotions are as consistent as possible. However, in the practical application of speech conversion, it is difficult to obtain a large number of parallel corpora, or even unsatisfactory. In addition, the accuracy of speech feature parameter vector alignment during training has also become a constraint on the performance of speech conversion systems. Regardless of the generality or practicability of speech conversion systems, the research on speech conversion methods under the condition of non-parallel texts has great practical significance and application value.

At present, there are two main methods of speech conversion under the condition of non-parallel text, the method based on speech clustering and the method based on parameter adaptation. The method based on speech clustering is to convert non-parallel texts into parallel texts under certain conditions by measuring the distance between speech frames or selecting corresponding speech units under the guidance of phoneme information. The principle of this method is simple, but the content of speech and text should be pre-extracted, and the result of pre-extraction will directly affect the conversion quality of speech. The method based on parameter adaptation is to use the speaker normalization or self-adaptation method in speech recognition to process the parameters of the conversion model, and its essence is to convert the pre-established model to the model based on the target speaker. This method can reasonably utilize the pre-stored speaker information, but usually the adaptive process will cause the smoothing of the spectrum, resulting in weak speaker personality information in the converted speech.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a speech conversion method that can adaptively determine the GMM mixing degree according to the difference of target speakers under the condition of non-parallel text, so as to enhance the speaker's personality characteristics in the converted speech. Also improve the quality of the converted speech.

The present invention adopts the following technical solutions for solving the above-mentioned technical problems:

The present invention proposes a speech conversion method under the condition of non-parallel text based on adaptive Gaussian clustering, including a training phase and a conversion phase, wherein the training phase includes the following steps:

Step 1, input the non-parallel training corpus of the source speaker and the target speaker;

Step 2: Use the AHOcoder speech analysis model to extract the MFCC feature parameter X of the non-parallel training corpus of the source speaker, the MFCC feature parameter Y of the non-parallel training corpus of the target speaker, and the source voice fundamental frequency log f _0X and the target voice base frequency. frequency log f _0Y ;

Step 3, for the MFCC feature parameters X, Y in step 2, perform unit selection and channel length normalization combined with voice feature parameter alignment and dynamic time regularization, thereby converting non-parallel corpus into parallel corpus;

Step 4, using the expectation maximization EM algorithm to perform the adaptive mixture Gaussian model AGMM training, after the AGMM training ends, the posterior conditional probability matrix P(X|λ) is obtained, and the AGMM parameter λ is saved;

Step 5, use the source speech feature parameter X and target speech feature parameter Y obtained in step 3, use the posterior conditional probability matrix P(X|λ) in step 4 to perform bilinear frequency bending BLFW+amplitude adjustment AS training, obtain Frequency bending factor α(x, λ) and amplitude adjustment factor s(x, λ) to construct BLFW+AS conversion function; use the mean and variance of logarithmic fundamental frequency to establish source speech fundamental frequency log f _0X and target speech fundamental frequency Fundamental frequency conversion function between frequencies log f _0Y ;

The conversion phase includes the following steps:

Step 6, input the source speaker voice to be converted;

Step 7, use the AHOcoder speech analysis model to extract the MFCC feature parameter X' and the logarithmic fundamental frequency log f _0X' of the source speaker's speech;

Step 8, use the parameter λ obtained during the AGMM training in step 4 to obtain the posterior conditional probability matrix P'(X|λ);

Step 9, use the BLFW+AS conversion function obtained in step 5 to obtain the converted MFCC characteristic parameter Y';

Step 10, using the fundamental frequency conversion function obtained in step 5 to obtain the converted logarithmic fundamental frequency log f _0Y' from the logarithmic fundamental frequency log f _0X ';

Step 11, use the AHOdecoder speech synthesis model to synthesize the converted MFCC feature parameter Y' and the logarithmic fundamental frequency log f _0Y' to obtain the converted speech.

Further, in the voice conversion method proposed by the present invention, the specific process of step 3 is as follows:

3-1) adopt the bilinear frequency bending method to carry out the channel length normalization processing to the source speech MFCC characteristic parameter;

3-2) For the given N source speech MFCC feature parameter vectors {X _k }, the N target speech feature parameter vectors {Y _k } are dynamically found by formula (1), so that the distance cost function value C({ Y _k }) is the smallest;

C({Y _k })=C ₁ ({Y _k })+C ₂ ({Y _k }) (1)

where C ₁ ({Y _k }) and C ₂ ({Y _k }) are respectively represented by the following equations:

Among them, the D(X _k , Y _k ) function represents the spectral distance between the source speech and the target speech feature parameter vector, the parameter γ represents the balance coefficient between the accuracy of the feature parameter frame alignment and the continuity between frames, and has 0≤γ≤1; C ₁ ({Y _k }) represents the spectral distance cost function between the source speech feature parameter vector and the target speech feature parameter vector, and C ₂ ({Y _k }) represents the unit selection The spectral distance cost function between the target speech feature parameter vectors;

即：3-3) By performing multiple linear regression analysis on formula (1), the target speech feature parameter sequence set aligned with the source speech feature parameter vector is obtained

which is:

Through the above steps, the non-parallel MFCC feature parameters X and Y are transformed into parallel corpus.

Further, in the speech conversion method proposed by the present invention, for the solution of formula (4), the Viterbi search method is used to optimize the execution efficiency of the algorithm.

Further, in the speech conversion method proposed by the present invention, the training process of step 4 is as follows:

4-1) Set AGMM initial mixture number M, Gaussian component weight coefficient thresholds t ₁ , t ₂ , Euclidean distance threshold D and covariance threshold σ between feature parameter vectors;

4-2) Use K-means iterative algorithm to obtain the initial value of EM training;

4-3) Use the EM algorithm for iterative training; express the Gaussian mixture model GMM as follows:

M为高斯分量的个数，N(X,μ_i,Σ_i)表示高斯分量的P维联合高斯概率分布，表示如下：Among them, X is a P-dimensional speech feature parameter vector, P=39; P( _wi ) represents the weight coefficient of each Gaussian component, and there are

M is the number of Gaussian components, and N(X, μ _i , Σ _i ) represents the P-dimensional joint Gaussian probability distribution of the Gaussian components, expressed as follows:

where μ _i is the mean vector, ∑ _i is the covariance matrix, λ={P( _wi ), μ _i ,∑ _i }, λ is the model parameter of the GMM model, and the estimation of λ is realized by the maximum likelihood estimation method, For the speech feature parameter vector set X={x _n , n=1,2,...N} there are:

at this time:

λ=arg _λ max(P(X|λ)) (8)

Use the EM algorithm to solve formula (8), as the iterative conditions in the EM calculation process satisfy P(X|λ ^k )≥P(X|λ ^k-1 ),

K is the number of iterations up to the model parameter λ. The iterative formulas of the Gaussian component weight coefficient P( _wi ), the mean vector μ _i and the covariance matrix Σ _i in the iterative process are as follows:

4-4) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i , ∑ _i ) in the model obtained by training is less than t ₁ , and its nearest neighbor component N(P(w _j ), μ _j , If the Euclidean distance between Σ _i ) is less than the threshold D, they are merged:

At this time, the number of Gaussian components becomes M-1, and return to step 4-3) for the next training. If there are multiple Gaussian components that meet the merging conditions, select the Gaussian component with the minimum distance for merging;

4-5) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i , ∑ _i ) in the model obtained by training is greater than t ₂ , and the variance of at least one dimension in the covariance matrix is greater than σ, it is considered that This Gaussian component contains excess information and should be split:

Among them, E is a column vector of all 1s, and n is used to adjust the Gaussian distribution. After splitting, the number of Gaussian components becomes M+1. If there are multiple Gaussian components that meet the splitting conditions, the component with the largest weight coefficient is selected for splitting. Return to step 4-3) for the next training;

4-6) After the AGMM training ends, the posterior conditional probability matrix P(X|λ) is obtained, and λ is saved.

Further, the speech conversion method proposed by the present invention, the BLFW+AS conversion function constructed in step 5, is expressed as follows:

F(x)=W _α(x,λ) x+s(x,λ) (15)

Among them, M is the number of Gaussian components of the Gaussian mixture model in step 4, α(x, λ) represents the frequency bending factor, and s(x, λ) represents the amplitude adjustment factor.

Further, the voice conversion method proposed by the present invention establishes the conversion relationship between the source voice pitch frequency and the target voice pitch frequency in step 5:

where μ and σ ² represent the mean and variance of the logarithmic fundamental frequency log f ₀ , respectively.

Compared with the prior art, the present invention adopts the above technical scheme, and has the following technical effects:

1. The present invention realizes the voice conversion under the condition of non-parallel text, solves the problem that parallel corpus is difficult to obtain, and improves the generality and practicability of the voice conversion system.

2. The present invention uses the combination of AGMM and BLFW+AS to realize a voice conversion system, which can adaptively adjust the classification number of GMM according to the distribution of voice characteristic parameters of different speakers, and improve the voice while enhancing the similarity of voice personality. quality, to achieve high-quality voice conversion.

Description of drawings

FIG. 1 is a schematic diagram of the non-parallel text-to-speech conversion of the present invention.

Figure 2 is a flowchart of the adaptive Gaussian mixture model training.

Fig. 3 is a spectrum comparison diagram of the converted speech.

Detailed ways

Below in conjunction with accompanying drawing, the technical scheme of the present invention is described in further detail:

It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as herein, are not to be taken in an idealized or overly formal sense. explain.

The high-quality voice conversion method of the present invention is divided into two parts: the training part is used to obtain parameters and conversion functions required for voice conversion, and the conversion part is used to convert the source speaker's voice into the target speaker's voice.

As shown in Figure 1, the implementation steps of the training part:

Step 1. Input the non-parallel corpus of the source speaker and the target speaker. The non-parallel corpus is taken from the CMU_US_ARCTIC corpus, which was established by the Institute of Language Technology of Carnegie Mellon University. The voices in the corpus are composed of 5 males and 2 Female recording, each speaker recorded 1132 segments of speech ranging from 1 to 6 seconds.

Step 2, the present invention uses the AHOcoder speech analysis model to extract the Mel-Frequency Cepstral Coefficients (MFCC, MFCC) X, Y and the logarithmic fundamental frequency parameters logf _0X and log f _0Y of the source speaker and the target speaker respectively. Among them, AHOcoder is a high-performance speech analysis and synthesis tool built by the team of Daniel Erro, a scholar at the Aholab Signal Processing Laboratory in Bilbao, Spain;

In step 3, the MFCC parameters X and Y of the source and target speech in step 2 are combined with unit selection (UnitSelection) and channel length normalization (VTLN, Vocal Tract Length No6rmalization) to combine speech feature parameter alignment and dynamic time warping. (DTW, Dynamic Time Warping). The specific process of speech feature parameter alignment is as follows:

3-1) The bilinear frequency bending method is used to normalize the channel length of the source speech feature parameters, so that the formants of the source speech are close to the target speech, thereby increasing the accuracy of the unit selection of the target speech feature parameters.

3-2) For the given N source speech feature parameter vectors {X _k }, the N target speech feature parameter vectors {Y _k } can be dynamically found by formula (1), so that the distance cost function value C({ Y _k }) is the smallest. In the process of unit selection, two factors are considered: on the one hand, the spectral distance between the aligned source speech feature parameter vector and the target speech feature parameter vector is minimized to enhance the matching degree of phoneme information; on the other hand, it is guaranteed The selected target speech feature parameter vector has frame continuity, so that the phoneme information is more complete.

C({Y _k })=C ₁ ({Y _k })+C ₂ ({Y _k }) (1)

where C ₁ ({Y _k }) and C ₂ ({Y _k }) can be respectively represented by the following equations:

Wherein, the D(X _k , Y _k ) function represents the spectral distance between the source and target feature parameter vectors, and the present invention adopts the Euclidean distance as the distance measurement scale. The parameter γ represents the balance coefficient between the accuracy of the feature parameter frame alignment and the continuity between frames, and has 0≤γ≤1. C ₁ ({Y _k }) represents the spectral distance cost function between the feature parameter vector of the source speech and the feature parameter vector of the target speech, and C ₂ ({Y _k }) represents the feature of the target speech selected by the unit Spectral distance cost function between parameter vectors.

即：3-3) By performing multiple linear regression analysis on formula (1), a feature parameter sequence set aligned with the source speech feature parameter vector can be obtained

which is:

For the solution of formula (4), the Viterbi search method can be used to optimize the execution efficiency of the algorithm.

Through the above steps, the non-parallel MFCC parameters X, Y are transformed into parallel.

In step 4, an adaptive Gaussian mixture model (Adaption GMM, AGMM) is established, and the expectation maximization (EM, Expectation-Maximization) algorithm is used for training, and the K-means iteration method is used to obtain the initial value of EM training. The AGMM parameters λ, P(X|λ) are obtained through training.

As shown in Figure 2, to use the adaptive clustering algorithm to train AGMM parameters, first of all, it is necessary to comprehensively analyze the Euclidean distance between the weight coefficient, mean vector, covariance matrix and feature parameter vector of each Gaussian component, and dynamically adjust the Gaussian distance. degree of mixing. The training process is as follows:

4-1) Set AGMM initial mixture number M, Gaussian component weight coefficient thresholds t ₁ , t ₂ , Euclidean distance threshold D and covariance threshold σ between feature parameter vectors.

4-2) Use the K-means iterative algorithm to obtain the initial value of EM training.

4-3) Use the EM algorithm for iterative training.

The traditional Gaussian mixture model is represented as follows:

M为高斯分量的个数，N(X,μ_i,∑_i)表示高斯分量的P维联合高斯概率分布，表示如下：Among them, X is a P-dimensional speech feature parameter vector, P=39 is used in the present invention, P( _wi ) represents the weight coefficient of each Gaussian component, and there are

M is the number of Gaussian components, and N(X,μ _i ,∑ _i ) represents the P-dimensional joint Gaussian probability distribution of the Gaussian components, which is expressed as follows:

where μ _i is the mean vector, ∑ _i is the covariance matrix, λ={P( _wi ), μ _i ,∑ _i } is the model parameter of the GMM model, and the estimation of λ can be done by the maximum likelihood estimation method (ML , Maximum Likelihood), the purpose of maximum likelihood estimation is to maximize the conditional probability P(X|λ), for the speech feature parameter vector set X={x _n , n=1,2,...N} there are:

at this time:

λ=arg _λ max(P(X|λ)) (8)

The EM algorithm can be used to solve formula (8). As the iterative conditions in the EM calculation process satisfy P(X|λ ^k )≥P(X|λ ^k-1 ), K is the number of iterations until the model parameter λ. In the iterative process, the iterative formulas of the Gaussian component weight coefficient P( _wi ), the mean vector μ _i , and the covariance matrix Σ _i are as follows:

4-4) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i , Σ _i ) in the model obtained by training is less than t ₁ , and its nearest neighbor component N(P(w _j ), μ _j , If the Euclidean distance between Σ _i ) is less than the threshold D, it is considered that these two components contain less information and are similar in composition, and can be merged:

At this time, the number of Gaussian components becomes M-1, and the process returns to step (3) for the next training. If there are multiple Gaussian components that meet the merging conditions, select the Gaussian component with the smallest distance for merging.

4-5) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i ,∑ _i ) in the model obtained by training is greater than t ₂ , and there is at least one-dimensional variance in the covariance matrix (covariance matrix pair The element on the corner is the variance) is greater than σ, then the Gaussian component is considered to contain excess information and should be split:

Among them, E is a column vector of all 1s, and n is used to adjust the Gaussian distribution. After splitting, the number of Gaussian components becomes M+1. If there are multiple Gaussian components that meet the splitting conditions, the component with the largest weight coefficient is selected for splitting. Return to step (3) for the next training.

4-6) After the AGMM training ends, the posterior conditional probability matrix P(X|λ) is obtained, and λ is saved.

Step 5, using the source speech feature parameter X obtained in step 3 and the target speech feature parameter Y and the posterior conditional probability matrix P(X|λ) obtained in step 4 for training to obtain a frequency bending factor and an amplitude adjustment factor, Thus, the bilinear frequency bending (BLFW, Bilinear Frequency Warping) and amplitude adjustment (AS, AmplitudeScaling) speech conversion functions are constructed, which are expressed as follows:

F(x)=W _α(x,λ) x+s(x,λ) (15)

Establish the conversion relationship between the source speech pitch frequency and the target speech pitch frequency:

where μ,σ ² are used to represent the mean and variance of the logarithmic fundamental frequency log f ₀ .

As shown in Figure 1, the specific implementation steps of the conversion part:

Step 6, input the source speaker voice to be converted;

Step 7, use the AHOdecoder speech analysis model to extract the 39th-order MFCC feature parameter X ' of the source speaker and the logarithmic fundamental frequency parameter log f _{0X '} of the source speech;

Step 8: Use λ={P( _wi ), μ _i , Σ _i } obtained during AGMM training in step 4 and the feature parameter X′ extracted in step 7, and substitute it into formula (5) to obtain the posterior conditional probability matrix P'(X|λ);

Step 9, use the frequency bending factor α(x, λ) and the amplitude adjustment factor s(x, λ) obtained by BLFW+AS training in step 5 and the posterior conditional probability matrix P′(X|λ) obtained in step 8 ), after substituting into formulas (15), (16), (17) and (18) respectively, the MFCC feature parameter Y′ of the converted speech is obtained;

Step 10, utilize the source voice logarithmic fundamental frequency parameter log f _0X' obtained in step 7, substitute formula (19), obtain the logarithmic fundamental frequency parameter log f _0Y' of voice after conversion;

Step 11, use the AHOdecoder speech synthesis model to obtain the converted speech by taking Y' in step 9 and log f _0Y' in step 10 as input.

Further, as shown in Figure 3, the method of the present invention is compared with the spectrogram of the converted voice obtained by the INCA method, and the conversion direction is F1-M2 (female voice 1-male voice 2), further verifying that the present invention adopts Compared with the INCA method, the method has the advantage of higher spectral similarity. Among them, the INCA method is the literature (Erro D, Moreno A, BonafonteA.INCA algorithm for training voice conversion systems from nonparallelcorpora[J].IEEE Transactions on Audio,Speech,and Language Processing,2010,18(5):944-953. ) proposed in.

The above are only some embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (7)

1. a speech conversion method under the non-parallel text condition based on adaptive Gaussian clustering, is characterized in that, comprises training stage and conversion stage, and wherein said training stage comprises the steps: Step 1, input the non-parallel training corpus of the source speaker and the target speaker; Step 2: Use the AHOcoder speech analysis model to extract the MFCC feature parameter X of the non-parallel training corpus of the source speaker, the MFCC feature parameter Y of the non-parallel training corpus of the target speaker, and the source voice fundamental frequency log f _0X and the target voice base frequency. frequency log f _0Y ; Step 3, for the MFCC feature parameters X, Y in step 2, perform unit selection and channel length normalization combined with voice feature parameter alignment and dynamic time regularization, thereby converting non-parallel corpus into parallel corpus; Step 4, using the expectation maximization EM algorithm to perform the adaptive mixture Gaussian model AGMM training, after the AGMM training ends, the posterior conditional probability matrix P(X|λ) is obtained, and the AGMM parameter λ is saved; Step 5, use the source speech feature parameter X and target speech feature parameter Y obtained in step 3, use the posterior conditional probability matrix P(X|λ) in step 4 to perform bilinear frequency bending BLFW+amplitude adjustment AS training, obtain Frequency bending factor α(x, λ) and amplitude adjustment factor s(x, λ) to construct BLFW+AS conversion function; use the mean and variance of logarithmic fundamental frequency to establish source speech fundamental frequency log f _0X and target speech fundamental frequency Fundamental frequency conversion function between frequencies log f _0Y ; The conversion phase includes the following steps: Step 6, input the source speaker voice to be converted; Step 7, use the AHOcoder speech analysis model to extract the MFCC feature parameter X' and the logarithmic fundamental frequency log f _0X' of the source speaker's speech; Step 8, use the parameter λ obtained during the AGMM training in step 4 to obtain the posterior conditional probability matrix P'(X|λ); Step 9, use the BLFW+AS conversion function obtained in step 5 to obtain the converted MFCC characteristic parameter Y'; Step 10, using the fundamental frequency conversion function obtained in step 5 to obtain the converted logarithmic fundamental frequency log f _0Y' from the logarithmic fundamental frequency log f _0X '; Step 11, use the AHOdecoder speech synthesis model to synthesize the converted MFCC feature parameter Y' and the logarithmic fundamental frequency logf _0Y' to obtain the converted speech. 2. voice conversion method according to claim 1, is characterized in that, step 3 concrete process is as follows: 3-1) adopt the bilinear frequency bending method to carry out the channel length normalization processing to the source speech MFCC characteristic parameter; 3-2) For the given N source speech MFCC feature parameter vectors {X _k }, the N target speech feature parameter vectors {Y _k } are dynamically found by formula (1), so that the distance cost function value C({ Y _k }) is the smallest; C({Y _k })=C ₁ ({Y _k })+C ₂ ({Y _k }) (1) where C ₁ ({Y _k }) and C ₂ ({Y _k }) are respectively represented by the following equations:

Among them, the D(X _k , Y _k ) function represents the spectral distance between the source speech and target speech feature parameter vectors, and the D(Y _k , Y _k-1 ) function represents the distance between the target speech feature parameter vectors selected by the unit Spectral distance, parameter γ represents the balance coefficient between the accuracy of feature parameter frame alignment and inter-frame continuity, and 0≤γ≤1; C ₁ ({Y _k }) represents the source speech feature parameter vector sum The spectral distance cost function between the target speech feature parameter vectors, C ₂ ({Y _k }) represents the spectral distance cost function between the target speech feature parameter vectors selected by the unit; 3-3) By performing multiple linear regression analysis on formula (1), the target speech feature parameter sequence set aligned with the source speech feature parameter vector is obtained

which is:

Through the above steps, the MFCC feature parameters X and Y under the non-parallel corpus are transformed into the aligned feature parameter set under the similar parallel corpus. 3 . The speech conversion method according to claim 2 , wherein, for the solution of formula (4), a Viterbi search method is used to optimize the execution efficiency of the algorithm. 4 . 4. speech conversion method according to claim 1, is characterized in that, the training process of step 4 is as follows: 4-1) Set AGMM initial mixture number M, Gaussian component weight coefficient thresholds t ₁ , t ₂ , Euclidean distance threshold D and covariance threshold σ between feature parameter vectors; 4-2) Use K-means iterative algorithm to obtain the initial value of EM training; 4-3) Use the EM algorithm for iterative training; express the Gaussian mixture model GMM as follows: Among them, X is a P-dimensional speech feature parameter vector, P( _wi ) represents the weight coefficient of each Gaussian component, and there are M is the number of Gaussian components, and N(X, μ _i , Σ _i ) represents the P-dimensional joint Gaussian probability distribution of the Gaussian components, expressed as follows:

where μ _i is the mean vector, ∑ _i is the covariance matrix, λ={P( _wi ), μ _i ,∑ _i }, λ is the model parameter of the GMM model, and the estimation of λ is realized by the maximum likelihood estimation method, For the speech feature parameter vector set X={x _n , n=1,2,...N} there are:

at this time: λ=arg _λ max(P(X|λ)) (8) Use the EM algorithm to solve formula (8), as the iterative conditions in the EM calculation process satisfy P(X|λ ^k )≥P(X|λ ^k-1 ), k is the number of iterations until the model parameter λ is obtained. The iterative formulas of the Gaussian component weight coefficient P( _wi ), the mean vector μ _i and the covariance matrix Σ _i in the iterative process are as follows:

4-4) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i , ∑ _i ) in the model obtained by training is less than t ₁ , and its nearest neighbor component N(P(w _j ), μ _j , If the Euclidean distance between Σ _i ) is less than the threshold D, they are merged:

At this time, the number of Gaussian components becomes M-1, and return to step 4-3) for the next training. If there are multiple Gaussian components that meet the merging conditions, select the Gaussian component with the minimum distance for merging; 4-5) If the weight coefficient of a Gaussian component N(P(w _i ), μ _i , ∑ _i ) in the model obtained by training is greater than t ₂ , and the variance of at least one dimension in the covariance matrix is greater than σ, it is considered that This Gaussian component contains excess information and should be split:

Among them, E is a column vector of all 1s, and n is used to adjust the Gaussian distribution. After splitting, the number of Gaussian components becomes M+1. If there are multiple Gaussian components that meet the splitting conditions, the component with the largest weight coefficient is selected for splitting. Return to step 4-3) for the next training; 4-6) After the AGMM training ends, the posterior conditional probability matrix P(X|λ) is obtained, and λ is saved. 5. The speech conversion method according to claim 4, characterized in that, P=39. 6. speech conversion method according to claim 1, is characterized in that, the BLFW+AS conversion function that constructs in step 5, is represented as follows: F(x)=W _α(x,λ) x+s(x,λ) (15)

Among them, M is the number of Gaussian components of the mixture Gaussian model in step 4,

Represents the posterior probability of the m-th Gaussian component of the speech feature vector x in the AGMM model λ, α _m and s _m represent the frequency bending factor and amplitude adjustment factor of the m-th Gaussian component in the AGMM model λ, respectively, α (x ,λ) represents the weighted combination of the frequency bending factors of all Gaussian components of the AGMM model λ, and s(x,λ) represents the weighted combination of the amplitude adjustment factors of all the Gaussian components of the AGMM model λ. 7. voice conversion method according to claim 1, is characterized in that, in step 5, establish the conversion relation between source voice pitch frequency and target voice pitch frequency:

where μ and σ ² represent the mean and variance of the logarithmic fundamental frequency log f ₀ , respectively.

Images