KR102613014B1 - Method for Automatic Severity Classification of Dysarthria - Google Patents

Abstract

The present invention relates to an automatic classification method for the severity of paralytic speech disorder that provides objective and consistent classification results based on 'extracted speech features'. The present invention relates to the Mel frequency used in clinical settings for automatic classification of the severity of paralyzed speech disorder. By applying the combination of cepstrum coefficient and the three speech function characteristics of sound quality, prosody, and pronunciation characteristics to machine learning to capture disability characteristics in a multifaceted manner, it objectively reflects speech intelligibility, which is an auditory-perceptual standard, and provides an intuitive interpretation of the results. can be provided.

Description

{Method for Automatic Severity Classification of Dysarthria}

The present invention relates to an automatic classification method for the severity of paralyzed speech disorder, and in particular, to an automatic classification method for the severity of paralyzed speech disorder that selects appropriate features from extracted speech features and provides objective and consistent classification results through a machine learning classifier. It's about.

Paralytic speech disorder is a term that refers to a speech motor disorder caused by paralysis and weakness of muscles related to speech production due to damage to the central nervous system and autonomic nervous system. Specifically, it occurs in a very wide range, including congenital brain damage (cerebral palsy), acquired brain damage (stroke), and degenerative neurological diseases due to aging (Parkinson's disease, Lou Gehrig's disease).

A speech rehabilitator determines the extent to which a speaker is having difficulties in communication situations by categorizing the severity of paralytic speech disorder and plans an intervention plan appropriate for the speaker's situation. In other words, assessing the severity of paralytic speech disorder can be seen as the beginning and milestone of speech therapy. The indicator mainly used in the diagnosis and treatment of paralytic speech disorder is specifically speech intelligibility. Speech intelligibility is assessed by speech rehabilitation specialists to determine the severity of a speaker's paralytic speech disorder and seek appropriate intervention measures accordingly.

In general, the severity assessment of paralyzed speech disorder based on speech intelligibility assessment by speech rehabilitation practitioners, which is mainly used in clinical practice, is an auditory-perceptual assessment that is not only subjective but also takes a lot of time and effort. Therefore, there is a request for an automatic classification technology for the severity of paralytic speech disorder that is similar to the speech intelligibility evaluation of an experienced speech rehabilitation worker. The specific requirement for this technology, which can be used as an assistive tool for speech rehabilitation practitioners, is to be able to provide objective and consistent results with a level of classification.

Recently, a technology that uses basic speech features such as MFCCs (Mel Frequency Cepstral Coefficients) and log filter banks as inputs to a deep neural network by applying deep learning methodology with a deep neural network (DNN) structure has been launched (Bhat). , C., & Strik, H. (2020). Automatic Assessment of Sentence-Level Dysarthria Intelligence Using BLSTM. IEEE Journal of Selected Topics in Signal Processing, 14(2), 322-330.). However, although the above methodology has the advantage of a simple data preparation process, it has the problem of not only requiring a large amount of data for each group, but also requiring a post-processing process for interpretation because the interpretation of the classification results is not intuitive.

Bhat, C., & Strik, H. (2020). Automatic Assessment of Sentence-Level Dysarthria Intelligence Using BLSTM. IEEE Journal of Selected Topics in Signal Processing, 14(2), 322-330.

The present invention aims to provide an automatic classification method for the severity of paralyzed speech disorder, which produces classification results at a level similar to the speech intelligibility evaluation of an experienced speech rehabilitator based on extracted speech features and does not require a separate post-processing process for interpretation.

The present invention is a method of automatically classifying the severity of paralytic speech disorder using an automatic classifier including a processor and memory, the method comprising: selecting a plurality of speech sentences (speech stimuli) from the paralytic speech disorder database corpus by using an automatic classifier including a processor and memory. ), selecting; Obtaining a voice file in which a speaker utters the plurality of speech sentences using an automatic classifier including a processor and memory; Mel Frequency Cepstrum Coefficients (MFCCs) for speech recognition, each of which extracts a plurality of sound quality features, prosody features, and pronunciation features, from the acquired voice file using an analysis toolkit by an automatic classifier including a processor and memory. step; selecting suitable features from the plurality of extracted sound quality features, prosody features, and pronunciation features by an automatic classifier including a processor and memory; And inputting suitable features selected from the mel frequency cepstrum coefficients, the sound quality features, the prosody features, and the pronunciation features into a trained machine learning classifier, by an automatic classifier including a processor and memory, and the sound quality features. The adaptive feature selected from the prosody feature is 'degree of voice breaks', the adaptive feature selected from the prosody feature is 'total length' and 'speech length', which are the speech rate, and the adaptive feature selected from the pronunciation feature is the consonant accuracy, which is phoneme accuracy. Percentage of Correct Consonants (PCC), Percentage of Correct Vowels (PCV), Percentage of Correct Total phonemes (PCT), and Vowel Space Area (VSA), which is the degree of vowel distortion, and vowel distortion. Vowel Articulatory Index (VAI), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio), and the machine learning classifier classifies speech into a plurality of severity levels based on the input features. It provides an automatic classification method for paralyzed speech disorder severity, which is a support vector machine (SVM) that classifies into grades.

The present invention also provides a method for automatically classifying the severity of paralytic speech disorder, wherein the severity levels are five levels: non-disability, mild, mild-moderate, moderate-severe, and severe.

The present invention also provides that the trained machine learning classifier, Mel frequency cepstrum coefficient, degree of voice breaks, total length, extracted from voice files of at least 10 non-disabled speakers and at least 70 paralytic speech impaired speakers Speech length, consonant accuracy, vowel accuracy, total phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and second formant slope are input into the support vector machine, and the voice is classified into a plurality of severity levels. Provides a method for automatically classifying the severity of paralytic speech disorder using data.

In addition, the present invention cross-validates the data classifying the voice into a plurality of severity levels through repeated measures analysis of variance (two-way mixed ANOVA), and the repeated measures analysis of variance is performed using the SPSS statistical analysis program, paralytic speech disorder. Provides an automatic severity classification method.

The present invention also provides a method for automatically classifying the severity of paralytic speech disorder, where the corpus is QoLT, a Korean paralytic speech disorder database.

The present invention also provides the above-mentioned utterance sentences: “On Chuseok, the whole family makes songpyeon together,” “I suddenly want to see my brother’s face in the United States,” and “Yesterday, the sky turned dark and rain poured down.” and “I got scolded by my mother because I fought with my younger brother.” We provide an automatic classification method for the severity of paralyzed speech disorder.

The present invention also provides a computer-readable storage medium storing a computer program for automatically classifying the severity of paralytic speech disorder, the storage medium comprising: code programmed to select a plurality of speech sentences (speech stimuli) from a corpus of paralytic speech disorder database. part time job; A code portion programmed to obtain a voice file in which a speaker utters the plurality of speech sentences; A code portion programmed to extract Mel Frequency Cepstrum Coefficients (MFCCs) for voice recognition, a plurality of sound quality features, prosody features, and pronunciation features from the acquired voice file using an analysis toolkit; a code portion programmed to select appropriate features from the plurality of extracted sound quality features, prosody features, and pronunciation features; and a code portion programmed to input fitting features selected from the mel frequency cepstrum coefficients, sound quality features, prosody features, and pronunciation features into a trained machine learning classifier, wherein the fit features selected from the sound quality features are 'speech breaks'. (voice breaks) degree', and the appropriate features selected from the prosody features are the speech rate, 'total length' and 'speech length', and the appropriate features selected from the pronunciation features are the phoneme accuracy, Percentage of Correct Consonants (PCC). ), Percentage of Correct Vowels (PCV), Percentage of Correct Total phonemes (PCT) and Vowel Space Area (VSA), which is the degree of vowel distortion, and Vowel Articulatory Index (VAI) ), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio), and the machine learning classifier is a support vector machine ( It provides a computer-readable storage medium, which is SVM: Support Vector Machine.

The present invention also provides that the severity level is five levels: non-disabled, mild, mild-moderate, moderate-moderate, and severe, and the trained machine learning classifier is used on at least 10 non-disabled speakers and at least 70 paralyzed speech-impaired speakers. Mel frequency cepstrum coefficient extracted from the voice file, degree of voice breaks, total length, speech length, consonant accuracy, vowel accuracy, overall phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and The second formant slope is input to the support vector machine to provide data that classifies the voice into a plurality of severity levels, and the data that classifies the voice into a plurality of severity levels is subjected to repeated measures analysis of variance (two-way mixed analysis). Cross-validation is performed through ANOVA), and the repeated measures analysis of variance is provided in a computer-readable storage medium, which is the SPSS statistical analysis program.

In the present invention, the corpus is QoLT, a Korean language paralysis database, and the utterance sentences are, “On Chuseok, the whole family makes songpyeon together.”, “I suddenly want to see my brother’s face in the United States.”, “ “Yesterday the sky turned dark and rain poured down.” and “I got scolded by my mom because I had a fight with my younger brother.” provides a computer-readable storage medium.

The present invention applies a combination of the Mel frequency cepstrum coefficient used in clinical practice and the three speech function features, voice quality, prosody, and pronunciation features, to automatic classification of the severity of paralytic speech disorder, to machine learning to capture the disability characteristics in a multifaceted manner, It can objectively reflect speech intelligibility, which is an auditory perceptual standard, and provide an intuitive interpretation of the results.

Figure 1 is a conceptual diagram of a method for automatically classifying the severity of paralytic speech disorder by extracting features from speech and inputting selected features from among the extracted features into a machine learning classifier, according to an embodiment of the present invention.
Figure 2 is a conceptual diagram of a model for finding a decision boundary that maximizes the margin of two groups to be distinguished in a support vector machine, according to an embodiment of the present invention.

Various aspects are disclosed with reference to the drawings. In the following description, for purposes of explanation, numerous specific details are set forth to facilitate a general understanding of one or more aspects. However, it will be appreciated that these aspects can be implemented without the specific details of each. The following description and accompanying drawings set forth in detail certain example aspects of one or more aspects. However, these aspects are illustrative and some of the various methods may be used in accordance with the principles of the various aspects and the written description is intended to encompass all such aspects and their equivalents.

Various aspects and features may be presented by a system that may include multiple devices, modules, etc. It should also be understood and appreciated that various systems may include additional devices, parts, components, etc. and/or may not include all of the devices, parts, components, etc. discussed in connection with the drawings.

As used herein, “embodiments,” “examples,” “aspects,” “examples,” etc. should not be construed to mean that any aspect or design described is better or advantageous over other aspects or designs. The terms 'system', 'server', terminal, etc. used below generally refer to computer-related entities and may mean, for example, hardware, a combination of hardware and software, or software.

Additionally, the term “or” is intended to mean an inclusive “or” and not an exclusive “or.” That is, unless otherwise specified or clear from context, “X utilizes A or B” is intended to mean one of the natural implicit substitutions. That is, either X uses A; X uses B; Or, if X uses both A and B, “X uses A or B” can be applied to either of the above cases. Additionally, the term “and/or” as used herein should be understood to refer to and include all possible combinations of one or more of the related listed items.

Additionally, the terms “comprise” and/or “comprising” mean that the feature, step, operation, module, and/or component is present, but one or more other features, steps, operations, modules, or components are present. It should be understood that this does not exclude the presence or addition of elements, and/or groups. In addition, although terms such as first and second may be used in this specification to describe various components, these components are not limited by these terms. That is, these terms are only used to distinguish between two or more components and should not be interpreted to imply order or priority. Additionally, unless otherwise specified or the context is clear to indicate a singular form, the singular terms herein and in the claims should generally be construed to mean “one or more.” Hereinafter, embodiments of the present invention will be described with reference to the attached drawings.

Figure 1 is a conceptual diagram of a method for automatically classifying the severity of paralytic speech disorder by extracting features from speech and inputting selected features from among the extracted features into a machine learning classifier, according to an embodiment of the present invention. The present invention is a method of automatically classifying the severity of paralytic speech disorder using an automatic classifier including a processor and memory, the method comprising: selecting a plurality of speech sentences (speech stimuli) from the paralytic speech disorder database corpus by using an automatic classifier including a processor and memory. ), selecting; Obtaining a voice file in which a speaker utters the plurality of speech sentences using an automatic classifier including a processor and memory; Mel Frequency Cepstrum Coefficients (MFCCs) for speech recognition, each of which extracts a plurality of sound quality features, prosody features, and pronunciation features, from the acquired voice file using an analysis toolkit by an automatic classifier including a processor and memory. step; selecting suitable features from the plurality of extracted sound quality features, prosody features, and pronunciation features by an automatic classifier including a processor and memory; and inputting suitable features selected from the mel frequency cepstrum coefficients, the sound quality features, the prosody features, and the pronunciation features into a trained machine learning classifier by an automatic classifier including a processor and memory.

In one embodiment of the present invention, the voice analysis toolkit that extracts features using the acquired voice file is an open source code and mainly uses the following types. In one implementation of the present invention, the voice file is analyzed through an input program using Python, etc., which includes the voice analysis toolkit.

- MFCCs extraction: librosa 0.8.1

- Sound quality feature extraction: Praat 6.1.16, Parselmouth 0.4.0

- Extraction of prosody (speech rate) features: Parselmouth 0.4.0

- Extraction of prosody (pitch) features: Praat 6.1.16

- Extraction of prosody (rhythm) features: Correlatore 2.3.4

- Pronunciation feature extraction: Kaldi and Praat 6.1.16.

The features extracted by the speech analysis toolkit are shown in Table 1.

[Table 1. Feature list]

In addition to MFCCs, which are the basic features of voice, sound quality features include jitter, shimmer, HNR, the number of voice breaks, and the degree of voice breaks. Among these, jitter, which represents the rate of frequency change, shimmer, which represents the rate of amplitude change, and HNR, which is the degree of noise in the voice, are This is an indicator mainly used when diagnosing voice disorders. The number and degree of voice breaks are characteristics used to check the speaker's ability to maintain voiced sounds. Voice breaks refer to cases where the length between consecutive pulses is longer than 17.86ms.

Prosodic features include overall length, speech length, speech rate, and articulation rate as speech speed, F0 mean value, standard deviation, minimum value, maximum value, median, 25th percentile, and 75th percentile as pitch, and %V, deltas, and varcos as rhythm. , rPVIs, and nPVIs. Since speakers with paralyzed speech have a slower speech rate than speakers without disabilities, the total length of the voice, the speech length excluding pauses in the total length, the speech rate calculated by dividing the total number of syllables by the total length, and the total number of syllables divided by the speech length. Values such as articulation speed are seen as characteristics. Speakers with paralyzed feet show different pitch patterns than speakers without disabilities, and F0 is the frequency of vocal folds per second, which perceptually corresponds to pitch, so various statistics of F0 are extracted as features. In addition, since paralyzed speech-impaired speakers have different rhythm patterns than non-disabled speakers, variables that have a strong correlation with rhythm elements are used as features. %V is the proportion of vowels in the entire speech, and deltas is the standard length of each vowel or consonant. Varcos is the deviation, where delta is normalized to the speech rate, rPVI is the interval between consecutive vowels or consonants, and nPVI is the rPVI normalized to the speech rate.

In addition, the pronunciation characteristics include phonemic accuracy, PCC, PCV, and PCT, and vowel distortion, Vowel Space Area (VSA), Formant Centralized Ratio (FCR), Vowel Articulatory Index (VAI), and F2-Ratio. The rationale for using phoneme accuracy as a pronunciation feature is that speakers with paralytic speech disorders show lower phoneme accuracy than speakers without disabilities. Specifically, PCC is the number of consonants with positive articulation out of the total number of target consonants, PCV is the number of vowels with positive articulation out of the total number of target vowels, and PCT is the number of phonemes with positive articulation among the total number of target phonemes. Additionally, vowel distortion is used because speakers with speech impairment have a narrower vowel area than speakers without speech impairment. VSA is the area of F1-F2 vowel squareness, FCR/VAI is an index of vowel formant centralization, and F2-Ratio is the ratio of the F2 values of plain and round vowels.

In one embodiment of the present invention, the appropriate feature selected from the sound quality feature is 'degree of voice breaks', and the appropriate feature selected from the prosody feature is 'total length' and 'speech length', which are the speech rate, and The suitable features selected from the pronunciation features are phoneme accuracy (Percentage of Correct Consonants (PCC)), vowel accuracy (Percentage of Correct Vowels (PCV)), total phoneme accuracy (Percentage of Correct Total phonemes (PCT)), and vowel distortion (vowel distortion). These are Vowel Space Area (VSA), Vowel Articulatory Index (VAI), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio). In one implementation of the present invention, the features extracted through analysis were classified as suitable features through the Extra Trees Classifier (ETC), a feature selection algorithm.

Using many features for classification does not necessarily lead to good results. Therefore, a process of summarizing features expected to be helpful in classification is necessary. This process is called the feature selection process. The feature selection algorithm used in one implementation of the present invention is an extra tree classifier, which is an ensemble algorithm that forms several decision trees and then collects classification results from each decision tree to reach a conclusion. Specifically, the learning of the extra tree classifier is carried out by randomly dividing the data and then selecting the best split so that data from different groups are not mixed within each group as much as possible.

Figure 2 is a conceptual diagram of a model for finding a decision boundary that maximizes the margin of two groups to be distinguished in a support vector machine, according to an embodiment of the present invention. A machine learning classifier learns patterns of suitable features selected as input values and classifies data based on the learned patterns. In one embodiment of the present invention, the machine learning classifier is a support vector machine (SVM) that classifies speech into a plurality of severity levels based on input features. SVM is one of the machine learning classifiers and is a model that finds the decision boundary that maximizes the margin, which is the distance between the two groups to be distinguished. The margin is calculated as the distance between the support vector, which is the closest data to the decision boundary. SVM, which has learned the optimal decision boundary that maximizes the margin, classifies which group the new input value belongs to based on this decision boundary.

The support vector machine classifies the severity of symptoms by group, and in one embodiment of the present invention, the severity grades are five grades: non-disability, mild, mild-moderate, moderate-moderate, and moderate. A machine learning classifier for such classification requires a training process. In one embodiment of the present invention, the trained machine learning classifier includes Mel frequency cepstrum coefficients, degree of voice breaks, and overall sound quality extracted from the voice files of at least 10 non-disabled speakers and at least 70 speakers with paralytic speech impairment. Length, speech length, consonant accuracy, vowel accuracy, total phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and second formant slope are input into the support vector machine to classify the voice into a plurality of severity levels. Provide classified data.

In one embodiment of the present invention, the data classifying the voice into a plurality of severity levels is cross-validated through repeated measures analysis of variance (two-way mixed ANOVA). Repeated measures analysis of variance is a statistical analysis method performed when checking whether independent variables show significant differences between groups (inter-subject factor) or when there is a relationship between independent variables (intra-subject factor). In other words, the purpose is to separate differences between and within individuals and identify their respective effects. In one embodiment of the present invention, the inter-subject factor was set as the severity group, and the intra-subject factor was set as speech function characteristics (sound quality, prosody, pronunciation). The program used for statistical analysis is the SPSS statistical analysis program.

In one embodiment of the present invention, the corpus is QoLT, a Korean language paralysis database, and using it, the classification method can be applied to the Korean language. In one embodiment of the present invention, the utterance sentences are “On Chuseok, the whole family makes songpyeon together.”, “I suddenly want to see my brother’s face in the United States.”, “Yesterday the sky became dark and it rained.” and “I got scolded by my mom because I fought with my younger brother.” These are sentences optimized for the analysis of speech paralysis created by the Department of Speech Pathology at Ewha Womans University.

The present invention is a computer-readable storage medium storing a computer program for automatically classifying the severity of paralytic speech disorder, the storage medium comprising: a code portion programmed to select a plurality of speech sentences (speech stimuli) from a paralytic speech disorder database corpus; A code portion programmed to obtain a voice file in which a speaker utters the plurality of speech sentences; A code portion programmed to extract Mel Frequency Cepstrum Coefficients (MFCCs) for voice recognition, a plurality of sound quality features, prosody features, and pronunciation features from the acquired voice file using an analysis toolkit; a code portion programmed to select appropriate features from the plurality of extracted sound quality features, prosody features, and pronunciation features; and a code portion programmed to input fit features selected from the mel frequency cepstrum coefficients, the tonal features, prosodic features, and pronunciation features into a trained machine learning classifier. In one embodiment of the present invention, the computer can obtain a voice file through a method such as recording a speaker's sentence utterance.

The appropriate feature selected from the sound quality features is 'degree of voice breaks', the appropriate feature selected from the prosody features is 'total length' and 'speech length', which are speech rate, and the appropriate feature selected from the pronunciation features is phoneme. Percentage of Correct Consonants (PCC), Percentage of Correct Vowels (PCV), Percentage of Correct Total phonemes (PCT), and Vowel Distortion: Vowel Space Area (VSA) ), Vowel Articulatory Index (VAI), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio), and the machine learning classifier classifies speech based on the input features. It is a Support Vector Machine (SVM) that classifies into multiple severity levels.

In one embodiment of the present invention, the computer running the storage medium can classify the severity into five levels: non-disability, mild, mild-moderate, moderate-moderate, and moderate. In addition, the trained machine learning classifier includes mel frequency cepstrum coefficient, degree of voice breaks, total length, speech length, and Consonant accuracy, vowel accuracy, overall phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and second formant slope are input into the support vector machine, and data is provided that classifies the voice into a plurality of severity levels. And the data classifying the voice into multiple severity levels can be cross-validated through repeated measures analysis of variance (two-way mixed ANOVA). In one embodiment of the present invention, the repeated measures analysis of variance is the SPSS statistical analysis program. In one embodiment of the present invention, the corpus is QoLT, a Korean language paralysis database, and the utterance sentences are “On Chuseok, the whole family makes songpyeon together,” and “I suddenly want to see my brother’s face in the United States.” , “Yesterday the sky turned dark and rain poured down.” and “I got into a fight with my younger brother and got scolded by my mom.”

According to software implementation, embodiments such as procedures and functions described in this specification may be implemented as separate software modules. Each of the software modules may perform one or more functions and operations described herein. Software code can be implemented as a software application written in an appropriate programming language. The software code may be stored in a management server and/or database and executed by an app.

Meanwhile, the various embodiments presented herein may be implemented as a method, device, or article manufactured using standard programming and/or engineering techniques. The term “article of manufacture” includes a computer program, carrier, or media accessible from any computer-readable device. For example, computer-readable media include magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips, etc.), optical disks (e.g., CDs, DVDs, etc.), smart cards, and flash memory devices. (e.g., EEPROM, card, stick, key drive, etc.), but is not limited to these. Additionally, various storage media presented herein include one or more devices and/or other machine-readable media for storing information. The term “machine-readable media” includes, but is not limited to, wireless channels and various other media capable of storing, retaining, and/or transmitting instruction(s) and/or data.

The description of the presented embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments without departing from the scope of the invention. Thus, the present invention is not limited to the embodiments presented herein, but is to be construed in the broadest scope consistent with the principles and novel features presented herein.

Claims (9)

A method of automatically classifying the severity of paralytic speech disorder using an automatic classifier including a processor and memory, the method includes:
Selecting a plurality of speech sentences (speech stimuli) from a paralytic speech disorder database corpus by an automatic classifier including a processor and memory;
Obtaining a voice file in which a speaker utters the plurality of speech sentences using an automatic classifier including a processor and memory;
Mel Frequency Cepstrum Coefficients (MFCCs) for speech recognition, each of which extracts a plurality of sound quality features, prosody features, and pronunciation features, from the voice file acquired using an analysis toolkit by an automatic classifier including a processor and memory. step;
selecting suitable features from the plurality of extracted sound quality features, prosody features, and pronunciation features by an automatic classifier including a processor and memory; and
Inputting suitable features selected from the mel frequency cepstrum coefficients, the sound quality features, the prosody features, and the pronunciation features into a trained machine learning classifier, by an automatic classifier including a processor and memory,
The appropriate feature selected from the sound quality features is 'degree of voice breaks', the appropriate feature selected from the prosody features is 'total length' and 'speech length', which are speech rate, and the appropriate feature selected from the pronunciation features is phoneme. Percentage of Correct Consonants (PCC), Percentage of Correct Vowels (PCV), Percentage of Correct Total phonemes (PCT), and Vowel Distortion: Vowel Space Area (VSA) ), Vowel Articulatory Index (VAI), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio),
The machine learning classifier is a support vector machine (SVM) that classifies speech into a plurality of severity levels based on input features.
Automatic classification method for paralytic speech disorder severity.
According to clause 1,
The severity level is five levels: non-disability, mild, mild-moderate, moderate-severe, and moderate.
Automatic classification method for paralytic speech disorder severity.
According to clause 1,
The trained machine learning classifier includes mel frequency cepstrum coefficient, degree of voice breaks, total length, speech length, and consonant accuracy extracted from the voice files of at least 10 non-disabled speakers and at least 70 speakers with paralysis and speech impairment. , vowel accuracy, overall phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and second formant slope are input into the support vector machine, and data is provided to classify the voice into a plurality of severity levels,
Automatic classification method for paralytic speech disorder severity.
According to clause 3,
The data classifying the voice into multiple severity levels was cross-validated through repeated measures analysis of variance (two-way mixed ANOVA),
The repeated measures analysis of variance is performed using the SPSS statistical analysis program,
Automatic classification method for paralytic speech disorder severity.
According to clause 1,
The above corpus is QoLT, a Korean language paralysis database,
Automatic classification method for paralytic speech disorder severity.
According to clause 5,
The above utterance sentence is,
“On Chuseok, the whole family makes songpyeon together.”, “Suddenly I miss my brother in America.”, “Yesterday the sky turned dark and rain poured down.” and “I got scolded by my mom for fighting with my younger brother.”
Automatic classification method for paralytic speech disorder severity.
A computer-readable storage medium that stores a computer program for automatically classifying the severity of paralytic speech disorder, the storage medium comprising:
A code portion programmed to select a plurality of speech stimuli (speech stimuli) from a paralytic speech disorder database corpus;
A code portion programmed to obtain a voice file in which a speaker utters the plurality of speech sentences;
A code portion programmed to extract Mel Frequency Cepstrum Coefficients (MFCCs) for speech recognition, a plurality of sound quality features, prosody features, and pronunciation features from the voice file acquired using an analysis toolkit;
a code portion programmed to select appropriate features from the plurality of extracted sound quality features, prosody features, and pronunciation features; and
Comprising a code portion programmed to input fit features selected from the mel frequency cepstrum coefficients, the sound quality features, the prosody features, and the pronunciation features into a trained machine learning classifier,
The appropriate feature selected from the sound quality features is 'degree of voice breaks', the appropriate feature selected from the prosody features is 'total length' and 'speech length', which are speech rate, and the appropriate feature selected from the pronunciation features is phoneme. Percentage of Correct Consonants (PCC), Percentage of Correct Vowels (PCV), Percentage of Correct Total phonemes (PCT), and Vowel Distortion: Vowel Space Area (VSA) ), Vowel Articulatory Index (VAI), Formant Centralized Ratio (FCR), and second formant slope (F2-Ratio),
The machine learning classifier is a support vector machine (SVM) that classifies speech into a plurality of severity levels based on input features.
Computer-readable storage media.
According to clause 7,
The severity level is five levels: non-disability, mild, mild-moderate, moderate-severe, and moderate.
The trained machine learning classifier includes mel frequency cepstrum coefficient, degree of voice breaks, total length, speech length, and consonant accuracy extracted from the voice files of at least 10 non-disabled speakers and at least 70 speakers with paralysis and speech impairment. , vowel accuracy, overall phoneme accuracy, vowel space area, vowel articulation index, formant centralization ratio, and second formant slope are input into the support vector machine, and data is provided that classifies the voice into a plurality of severity levels,
The data classifying the voice into multiple severity levels was cross-validated through repeated measures analysis of variance (two-way mixed ANOVA),
The repeated measures analysis of variance is performed using the SPSS statistical analysis program,
Computer-readable storage media.
According to clause 7,
The corpus is QoLT, a Korean language paralysis database,
The above utterance sentences are, “On Chuseok, the whole family makes songpyeon together,” “I suddenly miss my brother in the United States,” and “Yesterday, the sky turned dark and rain poured down.” and “I got scolded by my mom for fighting with my younger brother.”
Computer-readable storage media.