CN111583966A - Cross-database speech emotion recognition method and device based on joint distribution least square regression - Google Patents

Abstract

The invention discloses a cross-database speech emotion recognition method and device based on joint distribution least squares regression. The method includes: (1) acquiring a training database and a test database, wherein the training speech database includes several speech fragments and corresponding Speech emotion category label, the test database only contains a number of speech fragments to be recognized; (2) Use several acoustic low-dimensional descriptors to process and count the speech fragments, take each information obtained by the statistics as an emotional feature, and use it as an emotional feature. A plurality of emotional feature composition vectors are used as the feature vectors of the corresponding speech segments; (3) a least squares regression model based on joint distribution is established, and a sparse projection matrix is obtained by jointly training the training database and the test database; (4) For the speech segment to be recognized , obtain the feature vector according to step (2), and use the learned sparse projection matrix to obtain the corresponding voice emotion category label. The present invention can adapt to different environments and has higher accuracy.

Description

Method and device for cross-database speech emotion recognition based on joint distribution least squares regression

technical field

The present invention relates to speech emotion recognition, in particular to a cross-database speech emotion recognition method and device based on joint distribution least squares regression.

Background technique

The purpose of speech emotion recognition is to enable the machine to have enough intelligence to extract its emotional state (such as happiness, fear, sadness, etc.) from the speaker's speech. It is an important part of human-computer interaction and has huge research potential and development. prospect. For example, detecting the driver's mental state by combining the driver's voice, facial expressions and behavior information can remind the driver to concentrate on avoiding dangerous driving in time; detecting the interlocutor's voice emotion in the human-computer interaction can make the conversation smoother and take care of the interlocutor's behavior. Psychological, close to cognition; wearable devices can give more timely and intimate feedback based on the wearer's emotional state; at the same time, in various fields such as classroom teaching and teacher escort, speech emotion recognition is playing a more and more important role. increasingly important role.

Traditional speech emotion recognition is trained and tested on the same speech database, and the training and testing data follow the same distribution. In real life, the trained model must face different environments, and the sounding background will be mixed with various noises. Therefore, cross-database speech emotion recognition faces great challenges. How to make the trained model have good adaptability to different environments has become a problem that needs to be solved in academia and industry.

SUMMARY OF THE INVENTION

Purpose of the invention: The present invention provides a method and device for cross-database speech emotion recognition based on joint distribution least squares regression, aiming at the problems existing in the prior art. The present invention has good adaptability to different environments, and the recognition result is more accurate.

Technical solution: The cross-database speech emotion recognition method based on joint distribution least squares regression according to the present invention includes:

(1) obtain two voice databases, respectively as training database and test database, wherein, training voice database contains several voice fragments and corresponding voice emotion category labels, and test database only contains several voice fragments to be recognized;

(2) Use several acoustic low-dimensional descriptors to process and count the speech fragments, take each information obtained by the statistics as an emotional feature, and use a plurality of emotional features to form a vector as the feature vector of the corresponding speech fragment;

(3) Establish a least-squares regression model based on joint distribution, and use the training database of known labels and the test database of unknown labels to jointly train it to obtain a sparse projection matrix connecting the speech fragments and speech emotion category labels;

(4) For the speech segment to be recognized in the test database, the feature vector is obtained according to step (2), and the learned sparse projection matrix is used to obtain the corresponding speech emotion category label.

Further, step (2) specifically includes:

(2-1) For each speech segment, calculate its 16 acoustic low-dimensional descriptor values and corresponding increment parameters; the 16 acoustic low-dimensional descriptors are: time signal zero-crossing rate, frame energy root mean square , fundamental frequency, harmonic signal-to-noise ratio and Melton frequency cepstral coefficient 1-12;

(2-2) For each speech segment, 12 kinds of statistical functions are processed for its 16 acoustic low-dimensional descriptors respectively, and the 12 kinds of statistical functions are mean value, standard deviation, kurtosis, skewness, maximum value, minimum value, relative position, relative range, and two linear regression coefficients and their mean squared errors;

(2-3) Each information obtained by statistics is regarded as an emotional feature, and a plurality of emotional features are formed into a vector as the feature vector of the corresponding speech segment.

Further, the least squares regression model established in step (3) is:

表示找到使括号内式子最小的矩阵P，L_s∈R^c×n为训练数据库语音片段的语音情感类别标签向量，C为语音情感类别的类数，n为训练数据库语音片段的个数，X_s∈R^d×n为训练数据库语音片段的特征向量，d为特征向量的维数，P∈R^d×c为稀疏投影矩阵，P^T为P的转秩矩阵，

为Frobenius范数的平方，λ、μ为控制正则项的权衡系数，

X_t∈R^d×m为测试数据库语音片段的特征向量， m为测试数据库语音片段的段数，

分别为训练数据库、测试数据库中情感类别属于第c类的语音片段的集合，n_c、m_c分别为测试数据库中情感类别属于第c类的语音片段的个数，|| ||_2,1为2,1范数。In the formula,

means to find the matrix P that minimizes the expression in parentheses, L _s ∈ R ^c×n is the speech emotion category label vector of the speech segment in the training database, C is the number of speech emotion classes, n is the number of speech segments in the training database, X _s ∈R ^d×n is the feature vector of the speech segment in the training database, d is the dimension of the feature vector, P∈R ^d×c is the sparse projection matrix, P ^T is the rank transformation matrix of P,

is the square of the Frobenius norm, λ and μ are the trade-off coefficients that control the regular term,

X _t ∈ R ^d×m is the feature vector of the test database speech segment, m is the number of segments of the test database speech segment,

are the sets of speech clips whose emotion category belongs to the c-th category in the training database and the test database, respectively, n _c , m _c are the number of speech clips whose emotion category belongs to the c-th category in the test database, || || _2,1 is the 2,1 norm.

Further, described in step (3), the method of using the training database of the known label and the test database of the unknown label to jointly train it specifically includes:

(3-1) Convert the least squares regression model to:

s.t.P=Q

(3-2) Estimate the pseudo-label matrix formed by the pseudo-labels of the speech emotion categories corresponding to all speech segments in the test database through the transformed least squares regression model

统计得到

和m_c，进而计算得到

(3-3) According to the pseudo-label matrix

Statistics get

and m _c , and then calculate

对转换后的最小二乘回归模型利用增广拉格朗日乘子法进行求解，得到投影矩阵估计值

(3-4) Based on

The transformed least squares regression model is solved by the augmented Lagrange multiplier method to obtain the estimated value of the projection matrix

采用下式对伪标签矩阵

进行更新：(3-5) Estimated value according to projection matrix

The pseudo-label matrix is

To update:

表示中间辅助变量，

为

第i列第j行的元素，

表示求取第i列元素值最大的一行的行数j，

是伪标签矩阵

第i列第k行的元素；In the formula,

represents an intermediate auxiliary variable,

for

the element in column i and row j,

Represents the row number j of the row with the largest element value in the i-th column,

is the pseudo-label matrix

the element in column i and row k;

返回执行步骤(3-3)，直至达到预设的循环次数后，将循环结束后得到的的投影矩阵估计值

作为学习得到的投影矩阵P。(3-6) Using the updated pseudo-label matrix

Return to step (3-3), until the preset number of cycles is reached, the estimated value of the projection matrix obtained after the cycle ends

as the learned projection matrix P.

Further, step (3-2) specifically includes:

(3-2-1) Obtain the initial value of the estimated value of the projection matrix by using the formula of the transformed least squares regression model without adding a regular term

采用下式得到伪标签矩阵的初始值：(3-2-2) According to the initial value of the projection matrix

Use the following formula to get the initial value of the pseudo-label matrix:

表示中间辅助变量，

是伪标签矩阵的初始值

第i列第k行的元素。进一步的，步骤(3-4)具体包括：In the formula,

represents an intermediate auxiliary variable,

is the initial value of the pseudo-label matrix

The element in column i and row k. Further, step (3-4) specifically includes:

(3-4-1) Obtain the augmented Lagrangian equation of the least squares regression model:

In the formula, T is the Lagrange multiplier, k>0 is a regular parameter, and tr() represents the trace of the matrix;

(3-4-2) Keep P, T, and k unchanged, and update Q:

The part related to the variable Q in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

(3-4-3) Keep Q, T, and k unchanged, and update P:

The part related to the variable P in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

Pi is the _ith column vector of P, and Ti is the _ith column vector of T;

(3-4-4) Keep Q and P unchanged, and update T and k:

T=T+k(P-C)

k=min(ρk,k _max )

In the formula, k _max is the maximum value of the preset k, ρ is the scaling factor, ρ>1;

(3-4-5) Check for convergence:

Check whether ||PQ|| _∞ <ε is true, if not, return to step (3-4-2), if or the number of iterations is greater than the set value, the value of P at this time is used as the required sparse projection matrix , || || _∞ represents the largest element in the data, and ε represents the convergence threshold.

Further, the calculation method of the voice emotion category label of the test database described in step (4) is:

Calculated using the following formula:

表示中间辅助变量，j*表示待识别语音片段的语音情感类别标签。In the formula, P is the final projection matrix we have learned, X _t represents the feature vector set of the speech segment in the test database, that is, the feature vector set of the speech segment to be recognized,

represents the intermediate auxiliary variable, and j* represents the speech emotion category label of the speech segment to be recognized.

The cross-database speech emotion recognition device based on joint distribution least squares regression according to the present invention includes a processor and a computer program stored in a memory and running on the processor, and the processor implements the above method when executing the program .

Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the method and device for cross-database speech emotion recognition of the present invention are cross-database learning, therefore, they have good adaptability to different environments, and the recognition results are more accurate .

Description of drawings

FIG. 1 is a schematic flowchart of a cross-database speech emotion recognition method based on joint distribution least squares regression provided by the present invention.

Detailed ways

This embodiment provides a cross-database speech emotion recognition method based on joint distribution least squares regression, as shown in FIG. 1 , including the following steps:

(1) Acquire two speech databases as training database and test database, wherein the training speech database contains several speech fragments and corresponding speech emotion category labels, while the test database only contains several speech fragments to be recognized.

In this embodiment, we use three types of speech emotion databases commonly used in emotional speech recognition: Berlin, eNTERFACE and CAISA. Because the three types of databases contain different sentiment categories, the data are selected in the pairwise comparison. When comparing Berlin and eNTERFACE, we selected 375 pieces of data and 1077 pieces of data respectively, and the emotion categories were 5 categories (angry, scared, happy, disgusted, sad); when Berlin and CAISA were compared, we selected 408 pieces of data respectively. There are 1000 pieces of data and 1000 pieces of data, and the emotion category is 5 categories (angry, scared, happy, disgusted, sad); when eINTERFACE and CAISA are compared, we select 1072 pieces of data and 1000 pieces of data respectively, and the emotion category is 5 categories ( angry, scared, happy, disgusted, sad).

(2) Use several acoustic low-dimensional descriptors to process and count the speech fragments, take each information obtained by statistics as an emotional feature, and use multiple emotional features to form a vector as the feature vector of the corresponding speech fragment.

This step specifically includes:

(2-1) For each speech segment, calculate its 16 acoustic low-dimensional descriptor values and corresponding increment parameters; the 16 acoustic low-dimensional descriptors are: time signal zero-crossing rate, frame energy root mean square , fundamental frequency, harmonic signal-to-noise ratio and Melton frequency cepstral coefficients 1-12; the descriptor comes from the function set provided by INTERSPEECH 2009 Emotion Challenge;

(2-2) For each speech segment, use openSIMLE software to process 12 statistical functions for its 16 acoustic low-dimensional descriptors, and the 12 statistical functions are average, standard deviation, kurtosis, partial Degree, maximum value, minimum value, relative position, relative range, and two linear regression coefficients and their mean square errors;

(2-3) Take each information obtained by statistics as an emotion feature, and use 16×2×12=384 emotion features to form a vector as the feature vector of the corresponding speech segment.

(3) Establish a least squares regression model based on joint distribution, and use the training database of known labels and the test database of unknown labels to jointly train it, and obtain a sparse projection matrix connecting speech fragments and speech emotion category labels.

Among them, the established least squares regression model is:

表示找到使括号内式子最小的矩阵P，L_s∈R^c×n为训练数据库语音片段的语音情感类别标签向量，C为语音情感类别的类数，n为训练数据库语音片段的个数，X_s∈R^d×n为训练数据库语音片段的特征向量，d为特征向量的维数，P∈R^d×c为稀疏投影矩阵，P^T为P的转秩矩阵，

为Frobenius范数的平方，λ、μ为控制正则项的权衡系数，

X_t∈R^d×m为测试数据库语音片段的特征向量，m为测试数据库语音片段的段数，

分别为训练数据库、测试数据库中情感类别属于第c类的语音片段的集合，n_c、m_c分别为测试数据库中情感类别属于第c类的语音片段的个数，|| ||_2,1为2,1范数。In the formula,

means to find the matrix P that minimizes the expression in parentheses, L _s ∈ R ^c×n is the speech emotion category label vector of the speech segment in the training database, C is the number of speech emotion classes, n is the number of speech segments in the training database, X _s ∈R ^d×n is the feature vector of the speech segment in the training database, d is the dimension of the feature vector, P∈R ^d×c is the sparse projection matrix, P ^T is the rank transformation matrix of P,

is the square of the Frobenius norm, λ and μ are the trade-off coefficients that control the regular term,

X _t ∈R ^d×m is the feature vector of the test database speech segment, m is the number of segments of the test database speech segment,

are the sets of speech clips whose emotion category belongs to the c-th category in the training database and the test database, respectively, n _c , m _c are the number of speech clips whose emotion category belongs to the c-th category in the test database, || || _2,1 is the 2,1 norm.

Wherein, the method for jointly training the training database with the known label and the test database with the unknown label specifically includes:

(3-1) Convert the least squares regression model to:

s.t.P=Q

(3-2) Estimate the pseudo-label matrix formed by the pseudo-labels of the speech emotion categories corresponding to all speech segments in the test database through the transformed least squares regression model

统计得到

和m_c，进而计算得到

(3-3) According to the pseudo-label matrix

Statistics get

and m _c , and then calculate

对转换后的最小二乘回归模型利用增广拉格朗日乘子法进行求解，得到投影矩阵估计值

(3-4) Based on

The transformed least squares regression model is solved by the augmented Lagrange multiplier method to obtain the estimated value of the projection matrix

采用下式对伪标签矩阵

进行更新：(3-5) Estimated value according to projection matrix

The pseudo-label matrix is

To update:

表示中间辅助变量，

为

第i列第j行的元素，

表示求取第i列元素值最大的一行的行数j，

是伪标签矩阵

第i列第k行的元素；In the formula,

represents an intermediate auxiliary variable,

for

the element in column i and row j,

Represents the row number j of the row with the largest element value in the i-th column,

is the pseudo-label matrix

the element in column i and row k;

返回执行步骤(3-3)，直至达到预设的循环次数后，将循环结束后得到的的投影矩阵估计值

作为学习得到的投影矩阵P。(3-6) Using the updated pseudo-label matrix

Return to step (3-3), until the preset number of cycles is reached, the estimated value of the projection matrix obtained after the cycle ends

as the learned projection matrix P.

Further, step (3-2) specifically includes:

(3-2-1) Obtain the initial value of the estimated value of the projection matrix by using the formula of the transformed least squares regression model without adding a regular term

采用下式得到伪标签矩阵的初始值：(3-2-2) According to the initial value of the projection matrix

Use the following formula to get the initial value of the pseudo-label matrix:

表示中间辅助变量，

是伪标签矩阵的初始值

第i列第k行的元素。伪标签矩阵

的每一列只有其对应的类别那一行为1，其余行都为0。In the formula,

represents an intermediate auxiliary variable,

is the initial value of the pseudo-label matrix

The element in column i and row k. Pseudo-label matrix

Each column of is only 1 for its corresponding category, and the other rows are 0.

Step (3-4) specifically includes:

(3-4-1) Obtain the augmented Lagrangian equation of the least squares regression model:

In the formula, T is the Lagrange multiplier, k>0 is a regular parameter, and tr() represents the trace of the matrix;

(3-4-2) Keep P, T, and k unchanged, and update Q:

The part related to the variable Q in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

(3-4-3) Keep Q, T, and k unchanged, and update P:

The part related to the variable P in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

Pi is the _ith column vector of P, and Ti is the _ith column vector of T;

(3-4-4) Keep Q and P unchanged, and update T and k:

T=T+k(P-C)

k=min(ρk,k _max )

In the formula, k _max is the maximum value of the preset k, ρ is the scaling factor, ρ>1;

(3-4-5) Check for convergence:

Check whether ||PQ|| _∞ <ε is true, if not, return to step (3-4-2), if or the number of iterations is greater than the set value, the value of P at this time is used as the required sparse projection matrix , |||| _∞ means to find the largest element in the data,

ε represents the convergence threshold.

(4) For the speech segment to be recognized in the test database, the feature vector is obtained according to step (2), and the learned sparse projection matrix is used to obtain the corresponding speech emotion category label.

The specific method is to use the following formula to calculate the category label:

表示中间辅助变量，j^*表示待识别语音片段的语音情感类别标签。In the formula, P is the final projection matrix we have learned, X _t represents the feature vector set of the speech segment in the test database, that is, the feature vector set of the speech segment to be recognized,

represents the intermediate auxiliary variable, and j ^* represents the speech emotion category label of the speech segment to be recognized.

This embodiment also provides a cross-database speech emotion recognition device based on joint distributed least squares regression, including a processor and a computer program stored in a memory and executable on the processor, the processor executing the program implement the above method.

In order to verify the effectiveness of the present invention, we conduct experiments in pairs on the speech emotion database Berlin, eNTERFACE and CAISA database respectively. In each set of experiments, we use the two databases as the source domain and the target domain, where the source domain is used as the training set to provide training data and labels, and the target domain is used as the test set, providing only test data and no labels. In order to detect the recognition accuracy more effectively, we adopt two detection methods: Unweighted Average Recall (UAR) and Weighted Average Recall (WAR). Among them, UAR represents the number of correct predictions of each class divided by the number of participating tests for each class, and then averages the number of correct replacements for all classes; and WAR represents the number of all correct predictions divided by the number of all participating tests, regardless of The impact of each category quantity. Comprehensive consideration of UAR and WAR can effectively avoid the impact of unbalanced number of categories. As a comparative experiment, we selected several classic and efficient algorithms in subspace learning, namely: SVM, TCA, TKL, DaLSR, DoSL. The experimental results are shown in Table 1 below. Among them, the method is represented by the English abbreviation JDLSR in the table, the data set is the source domain/target domain, E, B, and C are the abbreviations of the eNTERFACE, Berlin, and CASIA data sets, respectively. The evaluation standard is UAR/WAR.

The experimental results show that, based on the micro-expression recognition method proposed by the present invention, a high cross-database micro-expression recognition rate is achieved.

Table 1

Claims (8)

1. a cross-database speech emotion recognition method based on joint distribution least squares regression, is characterized in that the method comprises: (1) two voice databases are obtained as training database and test database respectively, wherein, the training voice database contains several voice fragments and corresponding voice emotion category labels, and the test database only contains several voice fragments to be recognized; (2) Use several acoustic low-dimensional descriptors to process and count the speech fragments, take each information obtained by the statistics as an emotional feature, and use a plurality of emotional features to form a vector as the feature vector of the corresponding speech fragment; (3) Establish a least-squares regression model based on joint distribution, and use the training database of known labels and the test database of unknown labels to jointly train it to obtain a sparse projection matrix connecting the speech fragments and speech emotion category labels; (4) For the speech segment to be recognized in the test database, the feature vector is obtained according to step (2), and the learned sparse projection matrix is used to obtain the corresponding speech emotion category label. 2. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 1, is characterized in that: step (2) specifically comprises: (2-1) For each speech segment, calculate its 16 acoustic low-dimensional descriptor values and corresponding increment parameters; the 16 acoustic low-dimensional descriptors are: time signal zero-crossing rate, frame energy root mean square , fundamental frequency, harmonic signal-to-noise ratio and Melton frequency cepstral coefficient 1-12; (2-2) For each speech segment, 12 kinds of statistical functions are processed for its 16 acoustic low-dimensional descriptors respectively, and the 12 kinds of statistical functions are mean value, standard deviation, kurtosis, skewness, maximum value, minimum value, relative position, relative range, and two linear regression coefficients and their mean squared errors; (2-3) Each information obtained by statistics is regarded as an emotional feature, and a plurality of emotional features are formed into a vector as the feature vector of the corresponding speech segment. 3. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 1, is characterized in that: the least squares regression model that step (3) sets up is:

In the formula,

means to find the matrix P that minimizes the expression in parentheses, L _s ∈ R ^c×n is the speech emotion category label vector of the speech segment in the training database, C is the number of speech emotion classes, n is the number of speech segments in the training database, X _s ∈ R ^d×n is the feature vector of the speech segment in the training database, d is the dimension of the feature vector, P∈R ^d×c is the sparse projection matrix, P ^T is the rank transformation matrix of P,

is the square of the Frobenius norm, λ and μ are the trade-off coefficients that control the regular term,

X _t ∈R ^d×m is the feature vector of the test database speech segment, m is the number of segments of the test database speech segment,

are the sets of speech clips whose emotion category belongs to the c-th category in the training database and the test database, respectively, n _c , m _c are the number of speech clips whose emotion category belongs to the c-th category in the test database, || || _2,1 is the 2,1 norm. 4. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 3, is characterized in that: described in step (3) utilizes the training database of known label and the test database of unknown label to its The methods of joint training include: (3-1) Convert the least squares regression model to:

s.t.P=Q (3-2) Estimate the pseudo-label matrix formed by the pseudo-labels of the speech emotion categories corresponding to all speech segments in the test database through the transformed least squares regression model

(3-3) According to the pseudo-label matrix

Statistics get

and m _c , and then calculate

(3-4) Based on

The transformed least squares regression model is solved by the augmented Lagrange multiplier method to obtain the estimated value of the projection matrix

(3-5) Estimated value according to projection matrix

The pseudo-label matrix is

To update:

In the formula,

represents an intermediate auxiliary variable,

for

the element in column i and row j,

Represents the row number j of the row with the largest element value in the i-th column,

is the pseudo-label matrix

the element in column i and row k; (3-6) Using the updated pseudo-label matrix

Return to step (3-3), until the preset number of cycles is reached, the estimated value of the projection matrix obtained after the cycle ends

as the learned projection matrix P. 5. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 4, is characterized in that: step (3-2) specifically comprises: (3-2-1) Obtain the initial value of the estimated value of the projection matrix by using the formula of the transformed least squares regression model without adding a regular term

(3-2-2) According to the initial value of the projection matrix

Use the following formula to get the initial value of the pseudo-label matrix:

In the formula,

represents an intermediate auxiliary variable,

is the initial value of the pseudo-label matrix

The element in column i and row k. 6. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 4, is characterized in that: step (3-4) specifically comprises: (3-4-1) Obtain the augmented Lagrangian equation of the least squares regression model:

In the formula, T is the Lagrange multiplier, k>0 is a regular parameter, and tr() represents the trace of the matrix; (3-4-2) Keep P, T, and k unchanged, and update Q: The part related to the variable Q in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

(3-4-3) Keep Q, T, and k unchanged, and update P: The part related to the variable P in the augmented Lagrangian equation is proposed, and we get:

Solving the above formula, we get:

Pi is the _ith column vector of P, and Ti is the _ith column vector of T; (3-4-4) Keep Q and P unchanged, and update T and k: T=T+k(P-C) k=min(ρk,k _max ) In the formula, k _max is the maximum value of the preset k, ρ is the scaling factor, ρ>1; (3-4-5) Check for convergence: Check whether ||PQ|| _∞ <ε is true, if not, return to step (3-4-2), if or the number of iterations is greater than the set value, the value of P at this time is used as the required sparse projection matrix , || || _∞ represents the largest element in the data, and ε represents the convergence threshold. 7. the cross-database speech emotion recognition method based on joint distribution least squares regression according to claim 1, is characterized in that: the calculation method of test database speech emotion class label described in step (4) is: Calculated using the following formula:

In the formula, P is the projection matrix learned in step (3), X _t represents the feature vector set of the speech segment in the test database, that is, the feature vector set of the speech segment to be recognized,

represents the intermediate auxiliary variable, and j ^* represents the speech emotion category label of the speech segment to be recognized. 8. A cross-database speech emotion recognition device based on joint distribution least squares regression, comprising a processor and a computer program stored on a memory and running on the processor, characterized in that: the processor executes the program When implementing the method of any one of claims 1-6.

Images