Automatic dysarthria assessment method and system based on visual speech motor features

Abstract

The application discloses an automatic dysarthria evaluation method and system based on visual speech movement characteristics. Video stream data when expressing multiple unit sounds is acquired; after normalizing the human face image in each frame image, human face feature point calibration is performed; according to the time sequence of the video stream data and the calibrated human face feature points, the movement trajectory sequence of the lips and the lower jaw in the video stream data is extracted; according to the movement trajectory sequence, the distance from each detection point to the fixed point is calculated according to the preset detection point and the fixed point, and a distance sequence is formed; the formed distance sequence is input into the trained automatic dysarthria evaluation model, speech construction probability prediction is performed, and a dysarthria evaluation result is obtained. The application adopts more objective, comprehensive and detailed automatic evaluation indexes to assist the diagnosis and evaluation work of doctors, and provides a reliable and convenient method for subsequent speech therapists to formulate further speech rehabilitation treatment schemes, disease monitoring and the like for dysarthria patients.

Description

Technical Field

[0001] This invention relates to the field of articulation disorder technology, and in particular to an automated method and system for assessing articulation disorders based on visual speech motor characteristics. Background Technology

[0002] Articulation disorders refer to abnormalities or lesions in the muscular and nervous systems of organs involved in articulation, such as the lungs, vocal cords, soft palate, tongue, jaw, mouth, and lips, leading to an inability to accurately control their movements and thus manifesting as speech motor dysfunction. The causes of articulation disorders are diverse, commonly seen in conditions such as stroke, cerebral palsy, brain tumors, myasthenia gravis, cerebellar injury, Parkinson's disease, and multiple sclerosis. Patients often exhibit inaccurate pronunciation, slow speech duration, and disordered pitch and rhythm, which directly impacts their daily communication and quality of life. According to survey data, the number of people with speech disorders in China continues to rise. Therefore, a systematic analysis and research of articulation disorders is essential to improve the effectiveness of speech rehabilitation assessments for individuals with speech disorders.

[0003] Currently, in clinical practice, doctors mostly use subjective methods to assess the severity of dysarthria. Subjective evaluations are generally based on the therapist's own clinical experience or use standardized rating scales as auxiliary tools to assess the patient's speech intelligibility. However, these assessment methods heavily rely on the therapist's clinical experience and subjective judgment, leading to numerous influencing factors and problems such as unstable diagnostic results and low accuracy. With the rapid development of technologies such as deep learning and signal processing, some research progress has been made in the assessment of dysarthria. Current research on automated assessment of dysarthria mainly focuses on two aspects: feature set representation and extraction, and assessment model design and optimization. Therefore, this invention aims to utilize video data to accurately mine the correlation feature information within modalities, and build an objective assessment model of dysarthria based on visual speech motor features, providing reliable data and technical support for the medical diagnosis of dysarthria and the rehabilitation treatment of patients. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide an automated method and system for assessing articulation disorders based on visual speech motor features, thereby improving the diagnostic efficiency of clinical speech therapists. By selecting monosyllabic data from multiple vowels as raw data, and using the distance information of the lip and jaw movement trajectories during speech as feature representations, an automated articulation disorder assessment model fuses and splices the feature sets trained on six vowels at the decision level to ultimately obtain an assessment result of the severity of the articulation disorder. Based on the relevant characteristics of pathological speech in articulation disorders, this invention employs more objective, comprehensive, and detailed automated evaluation indicators to assist doctors in their diagnostic assessment work, providing a reliable and convenient method for speech therapists to develop further speech rehabilitation treatment plans and monitor the condition of patients with articulation disorders.

[0005] To achieve the above objectives, the present invention provides an automated method for assessing articulation disorders based on visual speech motor features, comprising the following steps:

[0006] S1. Obtain video stream data when multiple monotones are represented;

[0007] S2. Normalize the face images in each frame of the video stream data;

[0008] S3. Perform facial feature point calibration on the normalized face image;

[0009] S4. Based on the temporal sequence of the video stream data and the calibrated facial feature points, extract the motion trajectory sequence of the lips and jaw in the video stream data;

[0010] S5. In each frame of the image, based on the motion trajectory sequence, according to the preset detection points and fixed points, calculate the distance from each detection point to the fixed point to form a distance sequence;

[0011] S6. Input the formed distance sequence into the trained automated articulation disorder assessment model, which includes a feature extraction part and a classification assessment part;

[0012] The feature extraction section is used to form a feature sequence for each monophthong from the distance sequence according to the different selected vowels;

[0013] The classification and evaluation section is used to extract feature matrices from the input feature sequences, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder evaluation results.

[0014] Further preferably, in S1, face detection is performed on each frame of the acquired video stream data to determine the largest face detection box, and the face image of each frame of the video stream data is obtained by detecting the face image of each frame according to the largest face detection box.

[0015] Further preferably, in S3, facial feature point calibration is performed on the normalized face image; including: calibrating 68 feature points of the forehead, eyes, nose, lips, and jaw in the face image.

[0016] More preferably, in S5, the preset detection points and fixed points include: 13 points on the eyes and nose as fixed points; and 31 points on the lips and jaw as detection points.

[0017] More preferably, in S5, when calculating the distance from each detection point to the fixed point based on the motion trajectory sequence, the following formula (1) is used:

[0018] Formula (1)

[0019] in, Let x be the x-coordinate of the i-th feature point. Let be the ordinate of the i-th feature point. Let x be the x-coordinate of the i-th detection point. Let be the ordinate of the i-th detection point. ; Let x be the x-coordinate of the j-th fixed point. Let j be the ordinate of the fixed point. .

[0020] More preferably, in S4, the extraction of the motion trajectory sequence of the lips and jaw in the video stream data also includes detecting abnormal points in the motion trajectory sequence using the K-nearest neighbor-based outlier processing algorithm shown in the following formula (2);

[0021] Formula (2)

[0022] in, Is with The mean of the k nearest neighbors, Let the variance be k points. It is a constant; when an outlier occurs When, the average of the k nearest neighbors is taken. Replace the original value and fix the anomaly.

[0023] In a further preferred embodiment, in S6, the resulting distance sequence is uniformly framed according to the following formula (3);

[0024] Formula (3)

[0025] Where s represents the frame length to be standardized, and l represents the length of the original distance sequence. This represents the last frame of data in the distance sequence. Represents the original distance sequence. This represents the distance sequence after unifying the frame length, and then splicing them together.

[0026] The present invention also provides an automated articulation disorder assessment system based on visual speech motor features, comprising: a data acquisition module for acquiring video stream data when expressing multiple monophthongs;

[0027] The data processing module is used to normalize the face images in each frame of the video stream data;

[0028] For the normalized face image, facial feature points are labeled;

[0029] The data extraction module is used to extract the motion trajectory sequence of the lips and jaw from the video stream data based on the time sequence of the video stream data and the calibrated facial feature points; according to the preset detection points and fixed points, it calculates the distance from each detection point to the fixed point based on the motion trajectory sequence to form a distance sequence;

[0030] An automated articulation disorder assessment model includes a feature extraction part and a classification assessment part. The feature extraction part is used to form a feature sequence for each monophthong according to the selected vowels based on the distance sequence. The classification assessment part is used to extract a feature matrix from the input feature sequence, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder assessment result.

[0031] The present invention also provides an electronic device, including a processor and a memory, wherein the memory is used to store executable instructions of the processor; the processor is used to execute the stored instructions to implement the steps of the above-described automated articulation disorder assessment method based on visual speech motor features.

[0032] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the automated articulation disorder assessment method based on visual speech motor features as described above.

[0033] The automated articulation disorder assessment method and system based on visual speech motor characteristics disclosed in this application differs from traditional audio-based articulation disorder severity assessment methods and has at least the following advantages:

[0034] (1) This invention focuses on video data of articulation disorders, mines visual information in depth under this modality, improves the assessment effect of the severity of articulation disorders in the visual modality, and provides a reference for the research of multimodal articulation disorders.

[0035] (2) In the data input stage of the model, this invention designs a data organization method based on different monophthong syllables, so that the model can independently process the syllable data of each monophthong. Addressing the problem of articulation disorders in Chinese, this invention eliminates the problem of large intraclass differences caused by different speech content delivered by speakers.

[0036] (3) The back end of the model adopts a self-attention mechanism to extract the relative motion trajectory features of the lips and jaw. This method not only extracts motion features between local frames, but also provides global feature representation between frames and video sequences. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the automated articulation disorder assessment method based on visual speech motor features of the present invention.

[0038] Figure 2 This is a schematic diagram of 68 face detection points.

[0039] Figure 3 This is a matching diagram of real faces and detection points in a continuous video sequence.

[0040] Figure 4 This is a structural diagram of the proposed self-attention network based on speech motion distance. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Example 1

[0043] like Figure 1 As shown, one embodiment of the present invention provides an automated articulation disorder assessment method based on visual speech motor features, comprising the following steps:

[0044] S1. Acquire video stream data when expressing multiple monophthongs; S1 also includes performing face detection on each frame of the acquired video stream data, determining the largest face detection box, and detecting each frame of the video stream data according to the largest face detection box to acquire the face image of each frame.

[0045] Specifically, based on OpenCV and dlib tools, a face detector based on Histogram of Oriented Gradient (HOG) features is used to detect faces in each frame of the video and find the largest face detection box.

[0046] S2. Normalize the face images in each frame of the video stream data. Even for the same person, the detected face will vary in angle, brightness, etc. at different times. Therefore, the different face images are first normalized to a similar state to reduce the influence of useless, interfering, and redundant information such as background, brightness, and head posture. This process mainly includes face geometric normalization and grayscale normalization. Geometric normalization, also known as position calibration, corrects the angle tilt of the face caused by changes in posture. Grayscale normalization is used to compensate for face images obtained under different light intensities and light source directions to reduce the changes in image signals caused solely by changes in illumination.

[0047] S3. For the normalized face image, perform facial feature point calibration; such as... Figure 2 As shown, 68 feature points of the forehead, eyes, nose, lips, and jaw in the face image were calibrated; a regression tree model trained on the iBUG 300-W face feature database was used to calibrate the 68 feature points of the face.

[0048] S4. Based on the time sequence of the video stream data and the calibrated facial feature points, extract the motion trajectory sequence of the lips and jaw in the video stream data; more preferably, in S4, the extraction of the motion trajectory sequence of the lips and jaw in the video stream data also includes detecting abnormal points in the motion trajectory sequence using the K-nearest neighbor-based outlier processing algorithm shown in the following formula (2);

[0049] Formula (2)

[0050] in, Is with The mean of the k nearest neighbors, Let the variance be k points. It is a constant; when an outlier occurs When, the average of the k nearest neighbors is taken. Replace the original value and fix the anomaly.

[0051] S5. In each frame of the image, based on the motion trajectory sequence, according to the preset detection points and fixed points, calculate the distance from each detection point to the fixed point to form a distance sequence;

[0052] like Figure 2 As shown, in S5, the preset detection points and fixed points include: 13 points on the eyes and nose as fixed points; and 31 points on the lips and jaw as detection points.

[0053] In S5, when calculating the distance from each detection point to the fixed point based on the motion trajectory sequence, the following formula (1) is used:

[0054] Formula (1)

[0055] in, Let x be the x-coordinate of the i-th detection point. Let be the ordinate of the i-th detection point. ; Let x be the x-coordinate of the j-th fixed point. Let j be the ordinate of the fixed point. It should be noted that the distance from the detection point to the fixed point is calculated separately for each frame, with each frame independent of the previous one. In this application, the distance sequence formed by each frame is used as the feature input to the model, focusing only on the distance change between the detection point and the fixed point in the image, thus avoiding interference from the movement of the subject in consecutive frames.

[0056] S6. Input the formed distance sequence into the trained automated articulation disorder assessment model, which includes a feature extraction part and a classification assessment part;

[0057] The feature extraction section is used to form a feature sequence for each monophthong from the distance sequence according to the different selected vowels;

[0058] The classification and evaluation section is used to extract feature matrices from the input feature sequences, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder evaluation results.

[0059] In a further preferred embodiment, in S6, the resulting distance sequence is uniformly framed according to the following formula (3);

[0060] Formula (3)

[0061] Where s represents the frame length to be standardized, and l represents the length of the original distance sequence. This represents the last frame of data in the distance sequence. Represents the original distance sequence. This represents the distance sequence after unifying the frame length, while Concat represents splicing the original frame length sequence and the expanded frame length sequence.

[0062] The model was trained using cross-validation. The input features were preprocessed distance sequences from 31 points on the lips and jaw to 13 fixed reference points, with a size of 403. The automated articulation disorder assessment model consisted of two parts: a front-end feature extraction model and a back-end classification and assessment model.

[0063] like Figure 4 As shown, the first part is the feature extraction section, such as... Figure 4In the left branch, the preprocessed distance sequence of size B*6*T*403 is sequentially selected, and different vowel data of size B*T*403 are input into the front-end feature extraction model—Bidirectional Recurrent Neural Network (Bi-GRU). A random dropout strategy is used to extract the distance feature matrix, resulting in a feature sequence of size B*T*1024. , , ...... ).

[0064] Next is the backend classification and evaluation model, whose network structure is a multi-head self-attention mechanism. The input is the feature sequence B*T*1024 output by the frontend feature extraction model. For each monophthong syllable data, the processing steps are as follows.

[0065] 1) Convert the feature sequence of size B*T*1024 into a permutation of size T*B*1024.

[0066] 2) Input the feature sequence into the multi-head attention network to extract a feature matrix of size T*B*1024.

[0067] 3) Randomly dropout the feature matrix from 2) and add it to the feature matrix from 1).

[0068] 4) Normalize the feature matrix to obtain a feature matrix of size B*1024*T.

[0069] 5) Adaptive max pooling is used to obtain a feature matrix of size B*1024*1.

[0070] 6) Perform dimensionality reduction on the features and compress them into a feature matrix of size B*1024.

[0071] The feature matrices of size B*1024 for each monophthong syllable are concatenated to obtain a feature matrix of size B*6*1024. Finally, the matrix is ​​fed into a fully connected layer to obtain the prediction probability of each syllable. The final classification result is the average of the six prediction probabilities.

[0072] The present invention also provides an automated articulation disorder assessment system based on visual speech motor features, comprising: a data acquisition module for acquiring video stream data when expressing multiple monophthongs;

[0073] The data processing module is used to normalize the face images in each frame of the video stream data;

[0074] For the normalized face image, facial feature points are labeled;

[0075] The data extraction module is used to extract the motion trajectory sequence of the lips and jaw from the video stream data based on the time sequence of the video stream data and the calibrated facial feature points; according to the preset detection points and fixed points, it calculates the distance from each detection point to the fixed point based on the motion trajectory sequence to form a distance sequence;

[0076] An automated articulation disorder assessment model includes a feature extraction part and a classification assessment part. The feature extraction part is used to form a feature sequence for each monophthong according to the selected vowels based on the distance sequence. The classification assessment part is used to extract a feature matrix from the input feature sequence, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder assessment result.

[0077] The present invention also provides an electronic device, including a processor and a memory, wherein the memory is used to store executable instructions of the processor; the processor is used to execute the stored instructions to implement the steps of the above-described automated articulation disorder assessment method based on visual speech motor features.

[0078] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the automated articulation disorder assessment method based on visual speech motor features as described above.

[0079] Example 2

[0080] In this embodiment, data from 25 normal individuals and 25 patients were selected based on the severity of dysarthria and gender ratio. The automated dysarthria assessment method described in Example 1 was used for the experiment, with the specific parameters as follows:

[0081] The data used in this invention are a portion of video data from the MSDM corpus of articulation disorders. Monosyllabic data containing six monophthongs (a, o, e, i, u, v) were extracted as the original video data. Based on the severity of articulation disorders and the gender ratio, data from 25 normal individuals and 25 patients were selected. The specific data distribution is shown in Table 1. The Frenchay score is based on the scoring results of professional doctors.

[0082] Table 1 Dataset Distribution

[0083]

[0084] In S2, for each frame of the video stream, the center points of the two eyes are first determined using the average pixel coordinates of human eye feature points. Next, the angle between the line connecting the center points of the two eyes and the horizontal direction is calculated. Then, an affine transformation is used to rotate the face using the midpoint of the line connecting the center points of the two eyes as the origin to achieve geometric normalization. Finally, grayscale information is used to normalize the brightness. The final result is measured by the angle between the line connecting the center points of the eyes before and after registration and the horizontal line, and the angle between the face's central axis and the vertical line.

[0085] like Figure 3 In a video sequence, three consecutive frames are extracted. Speech motion trajectories are extracted from a real human face based on detection points. The points for the lips and jaw are: 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 48, 19, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67 (31 points in total). The points for the eyes and nose are: 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 39, 42, 45 (13 points in total). For each frame's detection point data, the distances between these 31 points and the 13 points are calculated, resulting in a distance sequence of size 31*13, or 403.

[0086] The processed distance sequence was used as input data. Six monosyllables containing different vowels were input into the front-end feature extraction model with an input size of B*T*403, where T is the value of 20 after uniform frame length in S4. After passing through a bidirectional recurrent neural network (Bi-GRU), the resulting feature sequence of size B*T*1024 was input into a multi-head self-attention network. After regularization and pooling operations, a feature matrix of size B*1024 was obtained. Finally, the six different feature matrices were concatenated, and two fully connected layers were used for classification prediction. The specific network structure and parameters are shown in Table 2.

[0087] Table 2. Network structure and parameter details of the automated articulation disorder assessment model

[0088]

[0089] During model training, a 5-fold cross-validation method was used, with each fold containing 5 normal individuals and 5 patients. The gender and severity of speech disorders among the speakers in each fold were generally balanced. The network training parameters had a learning rate of 3e8, a batch size of 32 samples in the training set, and used the Adam optimizer with 80 iterations.

[0090] The classification labels are four categories: mild, moderate, severe, and normal. The evaluation metrics for the classification results are precision, accuracy, recall, and F1 score, which are calculated using the following formulas:

[0091] Formula (4)

[0092] Formula (5)

[0093] recall Formula (6)

[0094] F1 Formula (7)

[0095] in, and Indicates true positive and false positive cases, TN and This indicates true negatives and false negatives.

[0096] Finally, the classification results of the 5-fold cross-validation were averaged to obtain the evaluation indicators. The results show that the method for automatically assessing the severity of articulation disorders based on visual-speech motor features, as described in this invention, achieves an accuracy of 79.65%, precision of 93.55%, recall of 65%, and F1 score of 76.7% at the syllable level. At the human level, the accuracy reaches 82%, precision of 94.44%, recall of 68%, and F1 score of 79.07%.

[0097] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An automated method for assessing articulation disorders based on visual-speech motor features, characterized in that, Includes the following steps: S1. Obtain video stream data when multiple monotones are represented; S2. Normalize the face images in each frame of the video stream data; S3. Perform facial feature point calibration on the normalized face image; S4. Based on the temporal sequence of the video stream data and the calibrated facial feature points, extract the motion trajectory sequence of the lips and jaw in the video stream data; S5. In each frame of the image, based on the motion trajectory sequence, according to the preset detection points and fixed points, calculate the distance from each detection point to the fixed point to form a distance sequence; S6. Input the formed distance sequence into the trained automated articulation disorder assessment model, which includes a feature extraction part and a classification assessment part; The feature extraction section is used to form a feature sequence corresponding to each monophthong from the distance sequence according to the selected vowels; including: The preprocessed distance sequence is then sequentially selected, and different vowel data are input into the front-end feature extraction model, a bidirectional recurrent neural network. A random dropout strategy is used to extract the distance feature matrix, resulting in a feature sequence of the first size. The classification and evaluation section is used to extract feature matrices from the input feature sequences, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder evaluation results. Using the feature sequence of the first size as input, the processing steps for each monophthong syllable data are as follows: 1) Transform the arrangement of the feature sequences of the first size; 2) Input the feature sequence of the first size into the multi-head attention network and extract it into the first feature matrix; 3) Randomly dropout the feature matrix from 2) and add it to the first feature matrix from 1); 4) Normalize the first characteristic matrix to obtain the second characteristic matrix; 5) Adaptive max pooling is used to obtain the third feature matrix; 6) Perform dimensionality reduction on the features, compressing them into a fourth feature matrix; The fourth feature matrix of each monophthong syllable data is concatenated to obtain the final feature matrix, which is then fed into a fully connected layer to obtain their respective prediction probabilities. The final classification result is the average of the six prediction probabilities.

2. The automated articulation disorder assessment method based on visual-speech motor features according to claim 1, characterized in that, In S1, face detection is performed on each frame of the acquired video stream data to determine the largest face detection box. Based on the largest face detection box, each frame of the video stream data is detected to obtain the face image of each frame.

3. The automated articulation disorder assessment method based on visual-speech motor features according to claim 1, characterized in that, In S3, facial feature points are calibrated on the normalized face image, including calibrating 68 feature points on the forehead, eyes, nose, lips, and jaw.

4. The automated articulation disorder assessment method based on visual-speech motor features according to claim 1, characterized in that, In S5, the preset detection points and fixed points include: 13 points on the eyes and nose as fixed points; and 31 points on the lips and jaw as detection points.

5. The automated articulation disorder assessment method based on visual-speech motor features according to claim 1, characterized in that, In S5, when calculating the distance from each detection point to the fixed point based on the motion trajectory sequence, the following formula (1) is used: Official (1) in, Let x be the x-coordinate of the i-th detection point. Let be the ordinate of the i-th detection point. ; Let x be the x-coordinate of the j-th fixed point. Let j be the ordinate of the fixed point. .

6. The automated articulation disorder assessment method based on visual speech motor features according to claim 1, characterized in that, In S4, the extraction of the motion trajectory sequence of the lips and jaw in the video stream data also includes the detection of abnormal points in the motion trajectory sequence using the K-nearest neighbor-based outlier processing algorithm shown in the following formula (2); Official (2) in, Is with The mean of the k nearest neighbors, Let the variance be k points. It is a constant; when an outlier occurs When, the average of the k nearest neighbors is taken. Replace the original value and fix the anomaly.

7. The automated articulation disorder assessment method based on visual-speech motor features according to claim 1, characterized in that, In S6, the resulting distance sequence is uniformly framed according to the following formula (3); Official (3) Where s represents the frame length to be standardized, and l represents the length of the original distance sequence. This represents the last frame of data in the distance sequence. Represents the original distance sequence. This represents the distance sequence after unifying the frame length, while Concat represents splicing the original frame length sequence and the expanded frame length sequence.

8. An automated articulation disorder assessment system based on visual-speech motor features, characterized in that, The implementation of the steps of the automated articulation disorder assessment method based on visual speech motor features according to any one of claims 1-7 includes: The data acquisition module is used to acquire video stream data when multiple monotones are represented; The data processing module is used to normalize the face images in each frame of the video stream data; For the normalized face image, facial feature points are labeled; The data extraction module is used to extract the motion trajectory sequence of the lips and jaw from the video stream data based on the time sequence of the video stream data and the calibrated facial feature points; according to the preset detection points and fixed points, it calculates the distance from each detection point to the fixed point based on the motion trajectory sequence to form a distance sequence; An automated articulation disorder assessment model includes a feature extraction part and a classification assessment part. The feature extraction part is used to form a feature sequence for each monophthong according to the selected vowels based on the distance sequence. The classification assessment part is used to extract a feature matrix from the input feature sequence, concatenate the feature matrices of each monophthong syllable data and perform articulation probability prediction to obtain the articulation disorder assessment result.

9. An electronic device, characterized in that, The device includes a processor and a memory, wherein the memory is used to store processor-executable instructions; the processor is used to execute the stored instructions to implement the steps of the automated articulation disorder assessment method based on visual speech motor features as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the automated articulation disorder assessment method based on visual speech motor features as described in any one of claims 1-7.

Images