JP4876207B2 - Cognitive impairment risk calculation device, cognitive impairment risk calculation system, and program - Google Patents

Description

  The present invention relates to a cognitive dysfunction risk calculation device, a cognitive dysfunction risk calculation system, and a program, and in particular, a cognitive dysfunction risk calculation device that calculates the risk of cognitive dysfunction based on voice data, The present invention relates to a cognitive impairment risk calculation system and program.

  Conventionally, screening for dementia includes HDS-R (revised Hasegawa simplified intelligence evaluation scale), MMSE (Mini-Mental State Examination), CDR (Clinical Dimensionia Rating), etc., and nerves such as fMRI, FDG-PET, and CSF biomarkers. It is widely used as well as physiologically based tests. These are implemented mainly in medical institutions by doctors who have received a certain amount of training or clinical psychologists.

  There is also known a dementia diagnosis support system that identifies a patient's dementia symptom level, generates a question and a correct answer according to the symptom level, and compares the patient's answer with the correct answer to determine correctness (patent) Reference 1).

JP 2007-282929 A

  However, the technique described in Patent Document 1 has a problem that, when performing repeated diagnosis, there is a possibility that the patient will remember the correct answer to the question, and the correct determination cannot be made.

  The present invention has been made to solve the above-described problems, and is a cognitive dysfunction risk calculating device, a cognitive dysfunction risk calculating system, and a program capable of accurately calculating the risk of cognitive dysfunction. The purpose is to provide.

  In order to achieve the above object, a cognitive impairment risk calculating apparatus according to a first aspect of the present invention is a cognitive impairment calculated for a plurality of types of prosodic feature quantities extracted from speech data and a speaker of the speech data. Based on a plurality of learning data including the risk level, feature quantity selection means for selecting a combination of the prosodic feature quantities that has the highest correlation with the risk level from the plurality of types of prosodic feature quantities; Weight determination means for determining a weight for each combination of the selected prosodic feature values based on the combination of the selected prosodic feature values and the degree of risk of each of the learning data, and input speech data From the feature quantity extracting means for extracting the plurality of types of prosodic feature quantities, and the selected prosodic feature among the prosodic feature quantities extracted by the feature quantity extracting means. And combinations of quantities, based on the weighting and determined by the weighting determining section is configured to include a a risk calculation means for calculating the risk of cognitive impairment.

  According to a second aspect of the present invention, there is provided a program for storing a plurality of pieces of learning data including a plurality of types of prosodic feature values extracted from speech data and a risk of cognitive impairment obtained for a speaker of the speech data. Based on the plurality of types of prosodic feature quantities, feature quantity selection means for selecting a combination of the prosodic feature quantities that has the highest correlation with the degree of risk, and the selected prosodic features of each of the plurality of learning data A weight determining means for determining a weight for each of the selected combinations of prosodic feature values based on a combination of amounts and the degree of risk, and a feature for extracting the plurality of types of prosodic feature values from input speech data A combination of the selected prosodic feature quantities among the prosodic feature quantities extracted by the quantity extracting means, the feature quantity extracting means, and the weight Only on the basis of the weighting and determined by the determining means, a program for functioning as a risk calculation means for calculating the risk of cognitive impairment.

  According to the first and second aspects of the present invention, the feature quantity selecting means extracts the plurality of types of prosodic feature quantities extracted from the speech data and the risk level of cognitive impairment determined for the speaker of the speech data. A combination of prosodic feature values having the highest correlation with the degree of risk is selected from a plurality of types of prosodic feature values based on the plurality of learning data included. The weight determination means determines the weight for each selected combination of prosodic feature values based on the selected combination of prosodic feature values and the degree of risk of each of the plurality of learning data.

  Then, a plurality of types of prosodic feature quantities are extracted from the input voice data by the feature quantity extracting means. Based on the combination of the prosodic feature quantities selected from the prosody feature quantities extracted by the feature quantity extracting means and the weighting determined by the weight determining means by the risk calculating means, the risk degree of cognitive dysfunction is calculated. calculate.

  In this way, the risk level of cognitive dysfunction is determined based on the combination of prosodic feature values having the highest correlation with the risk level of cognitive dysfunction and the weight determined for each of the prosodic feature value combinations. It is possible to calculate with high accuracy.

  A cognitive impairment risk calculating apparatus according to a third invention includes a plurality of types of prosodic feature quantities extracted from speech data, and a plurality of cognitive impairment risks determined for a speaker of the speech data. Analysis processing is performed on the plurality of types of prosodic feature values of the learning data to generate a plurality of types of composite variables obtained by combining the plurality of types of prosodic feature values, the plurality of learning data, Based on the plurality of types of generated synthetic variables, a combination variable selecting means for selecting a combination of the combined variables that has the highest correlation with the degree of risk from the plurality of types of combined variables, and the plurality of learning data Each of the selected combination variables based on the combination of the combination variables obtained for each of the plurality of learning data and the degree of risk of each of the plurality of learning data. Weighting determining means for determining a search, feature quantity extracting means for extracting the plurality of types of prosodic feature quantities from the input speech data, and the prosodic feature quantities extracted by the feature quantity extracting means And a risk degree calculating means for calculating the risk degree of cognitive impairment based on the combination of the composite variables and the weight determined by the weight determining means.

  According to a fourth aspect of the present invention, there is provided a program for storing a plurality of pieces of learning data including a plurality of types of prosodic feature quantities extracted from speech data and a risk level of cognitive impairment obtained for a speaker of the speech data. Analysis processing is performed on the plurality of types of prosodic feature values to generate a plurality of types of composite variables obtained by combining the plurality of types of prosodic feature values, the plurality of learning data, and the generated plurality of types Based on the synthetic variable of the type, the synthetic variable selection means for selecting the combination of the synthetic variable having the highest correlation with the degree of risk from the plurality of synthetic variables, each of the plurality of learning data is obtained. A weight for each of the selected combination of synthetic variables based on the combination of the synthetic variables and the risk of each of the plurality of learning data Weighting determining means for determining the input, feature quantity extracting means for extracting the plural types of prosodic feature quantities from the input speech data, and the composite variable obtained from the prosodic feature quantities extracted by the feature quantity extracting means And a weight calculating means for calculating the risk degree of cognitive impairment based on the weight determined by the weight determining means.

  According to the third and fourth aspects of the present invention, the synthetic variable generation means obtains a plurality of types of prosodic feature quantities extracted from the speech data, and the cognitive dysfunction risk calculated for the speaker of the speech data. An analysis process is performed on a plurality of types of prosodic feature values of a plurality of learning data included to generate a plurality of types of composite variables obtained by combining a plurality of types of prosodic feature values. Then, the combination variable selection means selects a combination of the combination variables having the highest correlation with the risk from the plurality of types of combination variables based on the plurality of learning data and the plurality of types of combination variables generated. . The weight determination means determines a weight for each selected combination of the composite variables based on the combination of the composite variables obtained for each of the plurality of learning data and the risk level of each of the plurality of learning data.

  Then, a plurality of types of prosodic feature quantities are extracted from the input voice data by the feature quantity extracting means. The risk level calculation means calculates the risk level of cognitive dysfunction based on the combination of the synthetic variables obtained from the prosodic feature values extracted by the feature value extraction means and the weights determined by the weight determination means.

  Thus, based on the combination of the prosodic feature variable that has the highest correlation with the risk of cognitive impairment and the weight determined for each combination of the combined variable, the cognitive impairment The degree of risk can be calculated with high accuracy.

  The above-mentioned plural types of prosodic feature quantities relate to feature quantities related to frequency components of speech, feature quantities related to speech formant structure, feature quantities related to speech volume, feature quantities related to speech speed, and reaction time until answering a question. At least one of the feature amounts can be included.

  The degree of risk of cognitive dysfunction in the learning data can be determined by the Hasegawa simple intelligence evaluation scale for the speaker.

  The feature quantity extraction means can extract a plurality of types of prosodic feature quantities from voice data input as an answer to a question.

  A cognitive dysfunction risk calculation system according to a fifth invention includes a plurality of types of prosodic feature quantities extracted from speech data, and a plurality of cognitive dysfunction risks determined for a speaker of the speech data. Based on learning data, feature quantity selection means for selecting a combination of the prosodic feature quantities having the highest correlation with the degree of risk from the plurality of types of prosodic feature quantities, and the selection of each of the plurality of learning data From the input speech data, a feature value selection device including weight determination means for determining a weight for each of the selected prosodic feature value combinations based on the prosodic feature value combination and the risk level, Feature quantity extraction means for extracting the plurality of types of prosodic feature quantities, and the selected prosody of the prosodic feature quantities extracted by the feature quantity extraction means A combination of symptoms amount, based on the weight determined by the weighting determining section is configured to include a a risk calculation device that includes a risk calculation means for calculating the risk of cognitive impairment.

  A cognitive dysfunction risk calculation system according to a sixth invention includes a plurality of types of prosodic feature quantities extracted from voice data and a plurality of cognitive dysfunction risk levels obtained for a speaker of the voice data. Performing principal component analysis on the plurality of types of prosodic feature values of the learning data to generate a plurality of types of composite variables obtained by combining the plurality of types of prosodic feature values, the plurality of learning data, Based on a plurality of types of synthesized variables generated, a synthesized variable selection means for selecting a combination of the synthesized variables having the highest correlation with the degree of risk from the plurality of types of synthesized variables, and the plurality of learning data Each of the selected combination variables based on the combination of the combination variables obtained for each of the plurality of learning data and the degree of risk of each of the plurality of learning data. A synthetic variable selection device including weight determination means for determining the weight to be determined, feature quantity extraction means for extracting the plurality of types of prosodic feature quantities from the input speech data, and the prosody extracted by the feature quantity extraction means A risk calculating device including a risk calculating means for calculating the risk of cognitive impairment based on the combination of the composite variables obtained from the feature amount and the weight determined by the weight determining means. It is configured.

  The program of the present invention can be provided by being stored in a recording medium.

  As described above, according to the cognitive dysfunction risk calculation device, the cognitive dysfunction risk calculation system, and the program of the present invention, the combination of prosodic feature quantities that has the highest correlation with the risk of cognitive dysfunction and Based on the weight determined for each combination of prosodic features, or a combination of synthetic variables of prosodic features and a combination of synthetic variables that have the highest correlation with the risk of cognitive impairment Based on the weights determined for each of the above, an effect is obtained that the risk of cognitive impairment can be calculated with high accuracy.

It is the schematic which shows the structure of the cognitive dysfunction risk calculation apparatus which concerns on the 1st Embodiment of this invention. (A) The graph which shows the frequency component extracted from audio | voice data, (B) The graph which shows a formant frequency. (A) The graph which shows the short time power extracted from audio | voice data, (B) The graph which shows the amplitude of audio | voice data. It is a flowchart which shows the content of the learning process routine of the computer which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the cognitive impairment disorder degree calculation process routine of the computer which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the cognitive dysfunction risk calculation apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows the example of the significant prosodic feature-value or synthetic | combination variable among the prosodic feature-values or synthetic variables selected by the method of 1st Embodiment and 2nd Embodiment. It is a figure which shows the correlation with the calculated risk and HDS-R score in the method of 1st Embodiment and 2nd Embodiment. It is a scatter diagram of the risk and the HDS-R score calculated by the method of the first embodiment. It is a scatter diagram of the risk and the HDS-R score calculated by the method of the second embodiment. It is the schematic which shows the structure of the cognitive dysfunction risk calculation system which concerns on the 3rd Embodiment of this invention. It is the schematic which shows the structure of the cognitive dysfunction risk calculation system which concerns on the 4th Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, a case where the present invention is applied to a cognitive impairment risk calculation apparatus that inputs the voice of a user who is a risk calculation target and calculates the risk of cognitive impairment will be described as an example. To do.

  As shown in FIG. 1, a cognitive dysfunction risk degree calculation device 10 according to an embodiment of the first invention includes a voice input unit 12 for inputting voice from a user, a speaker 13 for outputting voice, A computer 14 that calculates the risk of cognitive impairment based on the input voice data and causes the display device 16 to display it.

  The voice input unit 12 is composed of, for example, a microphone for inputting voice.

  The computer 14 includes a CPU, a ROM, a RAM, and an HDD. The HDD stores programs for a learning process routine and a cognitive function impairment risk calculation process routine, which will be described later.

  The computer 14 is functionally configured as follows. As shown in FIG. 1, the computer 14 includes a learning data storage unit 20, a feature amount selection unit 22, a weight determination unit 24, a question playback unit 25, a voice acquisition unit 26, a feature amount extraction unit 28, and a risk level calculation unit 30. It has.

  In the learning data storage unit 20, as learning data, data including a plurality of types of speech prosody features obtained from speech data and the cognitive dysfunction risk required for the speaker of the speech data is stored. It is stored in advance, and learning data for a large number of speakers is stored in advance.

  In this embodiment, in order to obtain learning data, voice data in which voice conversation is recorded when the revised Hasegawa simplified intelligence evaluation scale (HDS-R) is implemented as a verbal Q & A test is used. For example, we collect voices of answers to two questions about “registration” and “number recitation” from questions and answers, and in addition to this, three themes of hometown, childhood play, and school days Voice data was collected by collecting even the first phrase of any utterance voice from those chatting. A plurality of types of speech prosodic features obtained from the speech data obtained in this way and HDS-R scores representing the risk of cognitive impairment obtained by the test are collected as learning data, and learning data is collected. Stored in the storage unit 20.

  For example, the following 130 types of speech prosody features are used as the plurality of types of speech prosody features. When extracting these features, digital speech having 16 quantization bits and a sampling frequency of 44.1 [KHz] is used, the frame length in the short-time analysis is 23 [msec], the frame period is 11 [msec], and the window A Hamming window (1024 points) was used as a function.

  First, in order to reflect the pitch structure related to the pitch of speech, the following 53 types of speech prosodic features obtained from the fundamental frequency and n-order harmonic components having a frequency n times the fundamental frequency are used. It was. However, the amplitude of the time variation of the fundamental frequency is defined as the width between the maximum value and the minimum value when the upper 25% and lower 25% values of the data string of the basic frequency of one case are ignored.

F1-7. Amplitude of time change of fundamental frequency for t seconds immediately after the start of sound generation (t = 0.05, 0.10, ..., 0.35)
F8. Frequency centroid (weighted average of frequencies weighted by the power value of each harmonic component)
F10-48. Ratio of the total power value of the harmonic components from the fundamental to the nth order relative to the total power value of all harmonic components (n = 2, 3,..., 40)
F49. Ratio of sum of power values of odd-order harmonic components (including fundamental tone) and even-order harmonic components F50-53. Standard frequency standard deviation, average, maximum, minimum (see Fig. 2 (A))

  In order to reflect the formant structure representing the features of speech, 56 types of formant frequencies (see FIG. 2B) and formant bandwidths shown below are used as speech prosodic feature values.

F54-57. Standard deviation of nth formant frequency (n = 1, ..., 4)
F58-61. Average value of nth formant frequency (n = 1,..., 4)
F62-65. Maximum value of the nth formant frequency (n = 1,..., 4)
F66-69. Minimum value of the nth formant frequency (n = 1,..., 4)
F70-73. Median value of nth formant frequency (n = 1,..., 4)
F74-77. Difference between the maximum value and minimum value of the nth formant frequency (n = 1,..., 4)
F78-81. The slope of the linear approximation line of the nth formant frequency (n = 1,..., 4)
F82-85. Standard deviation of the nth formant bandwidth (n = 1,..., 4)
F86-89. Average value of nth formant bandwidth (n = 1,..., 4)
F90-93. Maximum value of nth formant bandwidth (n = 1,..., 4)
F94-97. Minimum value of nth formant bandwidth (n = 1,..., 4)
F98-101. Median of the nth formant bandwidth (n = 1,..., 4)
F102-105. Difference between the maximum value and the minimum value of the nth formant bandwidth (n = 1,..., 4)
F106-109. The slope of the linear approximation line of the nth formant bandwidth (n = 1,..., 4)

  Furthermore, in order to reflect the amplitude structure related to the volume of speech, the following 10 types of speech prosody features obtained from the short-time power and its envelope are used.

F110. Inclination of approximate straight line by linear least square method of power envelope F111-117. Median value of the derivative of the power envelope for t seconds immediately after the start of sounding (t = 0.05, 0.10, ..., 0.35)
F118-124. Ratio of maximum power value and power value at t seconds after the start of sound generation (t = 0.05, 0.10, ..., 0.35)
F125-128. Standard deviation, average, maximum, and minimum values of short-time power (see Fig. 3 (A))

  In order to reflect the time structure, the speed of speaking by the speaker and the reaction time until the question is answered are used as two types of prosodic feature quantities, as shown in FIG.

F129. Speech duration per mora F130. Reaction time to reply

  Next, the principle of selecting a combination of prosodic feature values will be described.

  The purpose of this embodiment is to discriminate between cognitive impairment (CI) and healthy (NL) by extracting and analyzing speech prosodic features from user speech data. However, when analyzing, if there are too many prosodic features extracted from speech data, the prosodic features may include prosodic features that are irrelevant to the discrimination of cognitive dysfunction. It is possible that the accuracy of construction and discrimination will be adversely affected. In addition, if there are too many prosodic feature quantities, the model becomes too complex and the calculation cost becomes high.

  Therefore, in the present embodiment, feature selection is performed on the 130 types of speech prosodic features described above. Feature selection methods include variable specification methods that specify appropriate variables based on scientific theory and prior knowledge, brute force methods that calculate combinations of all variables, and select the one that seems to be the best, and certain rules A sequential selection method that sequentially selects variables according to At present, speech features that are highly causally related to cognitive impairment in the elderly have not been identified, and there is no theory or prior knowledge useful for feature selection. In addition, calculating all combinations of the extracted prosodic feature values increases the calculation cost. Therefore, feature selection is performed using, for example, a forward stepwise method (FSW) as a commonly used sequential selection method. Note that the forward stepwise method is a method in which combinations are increased / decreased sequentially so as to obtain a better combination, and non-patent literature (Draper, N. and Smith, H .: Applied Regression Analysis (3rd edition) ), John Wiley & Sons (1998)), the detailed description is omitted.

  In the present embodiment, the feature amount selection unit 22 calculates the risk of cognitive impairment (HDS-R score) according to the forward stepwise method based on a large number of learning data stored in the learning data storage unit 20. A combination of speech prosodic feature values having the highest correlation is searched, and the searched combination of speech prosodic feature values is set as a combination of selected speech prosodic feature values.

  In addition, the weight determination unit 24 constructs a learning model that receives the selected combination of prosodic feature values and outputs the risk of cognitive impairment, and stores a large number of learning data stored in the learning data storage unit 20. Based on the selected combination of prosodic features and the risk of cognitive impairment (HDS-R score), for example, the combination of prosodic features is an explanatory variable and the HDS-R score is a target attribute A weighting for each selected prosodic feature is learned by performing multiple regression analysis. The learning model constructed here will be referred to as a model of cognitive dysfunction rating (SPCI) based on phonetic prosody (SPCI: speech prosthesis-based cognitive impulse rating).

  The question reproducing unit 25 reproduces preset question data from the speaker 25. For example, as questions data, questions in daily conversation such as “Where are you from?” Are used.

  The voice acquisition unit 26 acquires the user's voice data input by the voice input unit 12 as an answer to the question when the question is reproduced from the speaker 25.

  The feature quantity extraction unit 28 extracts the 130 types of speech prosodic feature quantities from the acquired user voice data.

  The degree-of-risk calculation unit 30 extracts a speech prosody feature amount extracted by the feature amount extraction unit 28 that corresponds to the combination of the prosodic feature amounts selected by the feature amount selection unit 22, and is determined by the weight determination unit 24. The extracted prosodic feature combinations are input to the learning model to which the weighting is applied, and the risk of cognitive impairment is calculated.

  Next, the operation of the cognitive dysfunction risk degree calculating device 10 according to the present embodiment will be described.

  First, an operator inputs a plurality of types of speech prosodic features and HDS-R obtained from learning speech data to the computer 14 as learning data, and stores them in the learning data storage unit 20.

  Then, a learning process routine shown in FIG. 4 is executed in the computer 14.

  First, learning data is read from the learning data storage unit 20 in step 100, and in step 102, the correlation with the HDS-R score is highest according to the forward stepwise method using the learning data read in step 100. A combination of speech prosodic feature values is searched, and the searched combination of speech prosodic feature values is set as a combination of selected speech prosodic feature values.

  In step 104, the learning data read in step 100 is used to learn and determine the weighting for each of the speech prosodic feature values selected in step 102 in the learning model.

  In the next step 106, the combination of the speech prosodic feature values selected in step 102 and the weight learned in step 104 are stored in a memory (not shown), and the learning processing routine is terminated.

  Each time the learning data in the learning data storage unit 20 is updated, such as when learning data is added, the above-described learning processing routine is executed again in the computer 14.

  When a risk level calculation instruction is input through the operation unit (not shown), the computer 14 executes a cognitive function disorder risk level calculation process routine shown in FIG.

  First, in step 118, question data is reproduced from the speaker 13, and in step 120, voice data input as a question answer via the voice input unit 12 is acquired. In step 122, 130 is obtained from the acquired voice data. Extracts phonetic prosody features of types.

  In step 124, a speech prosody feature value corresponding to the combination of the speech prosody feature values selected in the learning processing routine is selected from the speech prosody feature values extracted in step 122. In the next step 126, the combination of the phonetic prosody features selected in step 124 is input to the learning model to which the weighting learned in the learning processing routine is applied, and the risk level of cognitive dysfunction is calculated. In 128, the calculation result is displayed on the display device 16, and the cognitive impairment risk calculation processing routine is terminated.

  As described above, according to the cognitive dysfunction risk degree calculating device according to the embodiment of the first invention, the combination of the speech prosody feature amount having the highest correlation with the HDS-R score, and the speech prosody feature amount. Based on the weights learned for each of the combinations, the risk level of cognitive impairment can be calculated with high accuracy.

  In addition, when compared with the judgment method using the conventional revised Hasegawa simplified intelligence evaluation scale, it is not necessary to consider the contents of the language such as the correctness of the answer to the question and the correctness of the pronunciation, so objectively, The risk level of cognitive impairment can be determined. In addition, even when the risk level of cognitive dysfunction is repeatedly calculated, the risk level of cognitive dysfunction can be calculated with high accuracy.

  Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  In the second embodiment, the risk of cognitive impairment is calculated using a composite variable obtained by principal component analysis from a plurality of types of prosodic feature quantities. Is different.

  As shown in FIG. 6, the computer 214 of the cognitive dysfunction risk calculation apparatus 210 according to the second embodiment includes a learning data storage unit 20, a variable synthesis unit 221, a synthesis variable selection unit 222, a weight determination unit 224, A question reproduction unit 25, a voice acquisition unit 26, a feature amount extraction unit 28, a composite variable calculation unit 229, and a risk level calculation unit 230 are provided.

  The variable synthesis unit 221 performs variable synthesis by principal component analysis (PCA) based on, for example, 130 types of speech prosodic features of the learning data stored in the learning data storage unit 20, for example, 130 types of principal components ( Composite variable). In addition, the variable synthesis unit 221 converts each type of learning data from 130 types of speech prosody features to 130 types of synthesis variables.

  The synthetic variable selection unit 222 performs the cognitive impairment according to the forward stepwise method based on the synthetic variable of many learning data stored in the learning data storage unit 20 and the risk of cognitive impairment (HDS-R score). A combination of synthetic variables having the highest correlation with the degree of risk is searched, and the combination of searched synthetic variables is selected as a combination of synthetic variables to be selected.

  In addition, the weighting determination unit 224 receives the combination of the selected synthetic variables as input, constructs a learning model that outputs the risk of cognitive dysfunction, and selects a large number of learning data stored in the learning data storage unit 20. Based on the combination of the synthesized variables and the risk of cognitive impairment (HDS-R score), for example, a multiple regression analysis is performed with the combination of the synthetic variables as explanatory variables and the HDS-R score as a target attribute. Thus, the weighting for each synthetic variable in the learning model is learned.

  The synthesis variable calculation unit 229 calculates 130 types of synthesis variables generated by the variable synthesis unit 221 based on the 130 types of speech prosodic feature values extracted by the feature amount extraction unit 28.

  The risk degree calculation unit 230 takes out the combination variable calculated by the combination variable calculation unit 229 corresponding to the combination of the combination variables selected by the combination variable selection unit 222, and the weight determination unit 24 determines the weight. The combination of the extracted synthetic variables is input to the learning model, and the risk level of cognitive impairment is calculated.

  Next, a learning processing routine according to the second embodiment will be described.

  First, the learning data is read from the learning data storage unit 20, and the principal component analysis is performed based on the speech prosody feature value of the learning data to generate a synthesis variable, and the speech prosody feature value of each learning data is used as the synthesis variable. Convert. Then, using the learning data, in accordance with the forward stepwise method, a combination of synthetic variables having the highest correlation with the HDS-R score is searched, and the combination of searched synthetic variables is selected as a combination of synthetic variables to be selected. .

  Then, the learning data is used to learn and determine the weighting for each of the synthesis variables selected above in the learning model. Next, the combination of the synthesis variables selected above and the weighting learned above are stored in a memory (not shown), and the learning processing routine is terminated.

  Each time the learning data in the learning data storage unit 20 is updated, such as when learning data is added, the above-described learning processing routine is executed again in the computer 14.

  A cognitive impairment risk calculation processing routine according to the second embodiment will be described.

  First, question data is reproduced from the speaker 13, voice data input as a question answer via the voice input unit 12 is acquired, and 130 types of voice prosodic feature quantities are extracted from the acquired voice data. Then, 130 types of synthesis variables generated by the learning processing routine are calculated from the 130 types of extracted speech prosodic features.

  Then, a composite variable corresponding to the combination of the composite variables selected in the learning processing routine is selected from the composite variables calculated above. Next, the combination of the synthetic variables selected above is input to the learning model to which the weighting learned in the learning processing routine is applied, the risk level of cognitive impairment is calculated, and the calculation result is displayed on the display device 16 Then, the cognitive dysfunction risk degree calculation processing routine is terminated.

  Next, the result of calculating the risk of cognitive impairment using the techniques of the first and second embodiments will be described.

Here, a total of 319 audio data were collected from 115 elderly people (aged 38-99 years old, 32 men, 83 women), the risk of cognitive impairment was calculated, and the HDS-R score The correlation was calculated. 7 to 10, the calculation result obtained by the cognitive impairment risk calculation method described in the first embodiment is represented as SPCI _FSW-AIC, and the cognitive impairment disorder risk described in the second embodiment. The calculation result obtained by the calculation method is represented as SPCI _PCA-FSW-AIC .

  As shown in FIG. 7 and FIG. 8, in the method of the first embodiment, 19 prosodic feature quantities are selected, of which F129, F128, F118, F130, F57, F8, F101, F59, F110, F72, F69, and F73 were determined to be significant prosodic feature values. Thus, it was found that a significant prosodic feature value was selected by feature value selection using the AIC criterion.

The correlation R with the HDS-R score was 0.67, and the corrected determination coefficient (reliability) was R ² = 0.41. As shown in FIG. 9, it was found that there was a relatively significant correlation between the calculated risk of cognitive impairment and the HDS-R score.

  In the method of the second embodiment, 55 principal components (synthetic variables) obtained by appropriately synthesizing all 130 types of features are selected, and many synthetic variables are determined to be significant. It was shown that principal components having low eigenvalues such as PC77, PC115, and PC103 as well as principal components having high eigenvalues (for example, PC2, PC7, PC4) are useful for HDS-R score estimation.

Further, the correlation R with the HDS-R score is 0.77, and as shown in FIG. 10, it is suggested that there is a strong correlation between the calculated risk of cognitive impairment and the HDS-R score. It was done. Further, the corrected determination coefficient R ² = 0.50, which is a significant result as an applicability to the screening of cognitive dysfunction in speech analysis of elderly people based on prosodic features.

As described above, according to the cognitive dysfunction risk calculation apparatus according to the second embodiment, the combination of the speech prosodic feature value synthesis variable and the synthesis variable that have the highest correlation with the HDS-R score. Based on the weights learned for each of the combinations, the risk level of cognitive impairment can be calculated with high accuracy.
In the embodiment described above, the case where the principal component analysis is performed to generate the synthesis variable of the speech prosody feature amount is described as an example. However, the present invention is not limited to this. For example, the factor analysis is performed. Alternatively, a synthesis variable of the speech prosody feature value may be generated.

  Next, a third embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  In the third embodiment, the first point is that a device for selecting prosody features and learning for weighting from learning data and a device for calculating the risk of cognitive impairment are connected via a network. This is mainly different from the embodiment.

  As shown in FIG. 11, the cognitive dysfunction risk calculation system 310 according to the third embodiment includes a feature amount learning device 311 that performs prosody feature amount selection and weight learning from learning data, and input speech. A plurality of cognitive dysfunction risk calculating devices 312 that calculate the risk of cognitive dysfunction from the data, and the feature quantity learning device 311 and the plurality of cognitive dysfunction risk calculating devices 312 include the Internet, etc. Network 313.

  The feature quantity learning device 311 is constituted by a computer server, for example, and is functionally configured as follows. As shown in FIG. 11, the feature quantity learning device 311 includes a learning data storage unit 20, a feature quantity selection unit 22, a weight determination unit 24, and a parameter transmission unit 324.

  The parameter transmission unit 324 receives a parameter indicating the combination of the prosodic feature values selected by the feature value selection unit 22 and the weight of the learning model determined by the weight determination unit 24 via the network 313. To the degree calculation device 312.

  The cognitive function disorder risk calculation device 312 includes a voice input unit 12, a speaker 13, a computer 314, and a display device 16.

  The computer 314 includes a parameter reception unit 325, a question reproduction unit 25, a voice acquisition unit 26, a feature amount extraction unit 28, and a risk level calculation unit 30.

  The parameter receiving unit 325 receives the parameter transmitted from the feature amount learning device 311 and stores it in a memory (not shown).

  The risk level calculation unit 30 extracts speech prosodic feature values extracted by the feature value extraction unit 28 that correspond to the combination of prosodic feature values indicated by the parameters stored in the memory, and the parameters stored in the memory are extracted. The combination of the extracted prosodic feature quantities is input to a learning model to which the weighting shown is applied, and the risk of cognitive impairment is calculated.

  Next, the operation of the cognitive dysfunction risk calculation system 310 according to the present embodiment will be described.

  The operator inputs a plurality of types of prosodic features and HDS-R obtained from the learning speech data to the feature learning device 311 as learning data, and the learning data stored in the learning data storage unit 20 Each time it is added, the learning process routine described in the first embodiment is executed. Then, the feature amount learning device 311 transmits a parameter indicating the selected combination of the phonetic prosody feature amount and the learned weighting to the cognitive dysfunction risk calculating device 312 via the network 313.

  The cognitive dysfunction risk degree calculation device 312 receives the above parameters and stores them in a memory (not shown).

  In addition, when a risk level calculation instruction is input to the cognitive function disorder risk level calculation device 312, the computer 314 executes the cognitive function level risk level calculation processing routine described in the first embodiment. .

  In this way, the combination of speech prosody features that has the highest correlation with the HDS-R score and the weights learned for each combination of speech prosody features are learned by the feature value learning device, and the network By distributing the learning results to a plurality of cognitive impairment risk calculators, each cognitive impairment risk calculator uses the latest learning results to calculate the risk of cognitive impairment. be able to.

  Note that the above-described embodiment may be applied to the configuration of the above-described second embodiment. In this case, the cognitive dysfunction risk degree calculation system includes a synthetic variable learning apparatus that performs selection of a synthetic variable and learning of weighting from learning data, and a plurality of cognitive dysfunction risk risk calculation apparatuses connected by the network 313. The combined variable learning apparatus includes a learning data storage unit 20, a variable combining unit 221, a combined variable selecting unit 222, a weight determining unit 224, and a parameter transmitting unit 324. The computer of the cognitive dysfunction risk calculation apparatus includes a parameter reception unit 325, a question reproduction unit 25, a voice acquisition unit 26, a feature amount extraction unit 28, a synthetic variable calculation unit 229, and a risk calculation unit 30.

  Next, a fourth embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment and 3rd Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The fourth embodiment is mainly different from the third embodiment in that the input / output terminal and the cognitive dysfunction risk calculating device are connected via a network.

  As shown in FIG. 12, the cognitive dysfunction risk calculation system 410 according to the fourth embodiment includes a plurality of input / output terminals 412 and a cognitive dysfunction risk calculation device 414, and a plurality of input / output terminals. The terminal 412 and the cognitive dysfunction risk degree calculation device 414 are connected via a network 313.

  The input / output terminal 412 includes a voice input unit 12, a speaker 13, a display device 16, a question reproduction unit 25, a voice acquisition unit 26, and a communication unit 425, and inputs voice as an answer to a question reproduced from the speaker 13. When voice data is input from the user via the unit 12, the communication unit 425 transmits the voice data to the cognitive impairment risk calculating device 414 via the network 313.

  When the communication unit 424 receives the voice data, the cognitive dysfunction risk calculation device 414 calculates the risk of cognitive dysfunction based on the voice data, and the communication unit 424 sends the calculation result via the network 313. To the input / output terminal 412.

  When the communication unit 425 receives the calculation result, the input / output terminal 412 displays the calculated risk level of cognitive impairment on the display device 16.

  In addition, since it is the same as that of 1st Embodiment about the other structure and effect | action of the cognitive function disorder risk calculation apparatus 414, description is abbreviate | omitted.

  In this way, multiple input / output terminals and a cognitive dysfunction risk calculation device are connected via a network, and voice data and calculation results are transmitted / received to determine the risk of cognitive dysfunction at each input / output terminal. can do.

  Note that the above-described embodiment may be applied to the configuration of the above-described second embodiment. In this case, the cognitive dysfunction risk calculation system includes each input / output terminal 412 and a device having the same function as the cognitive dysfunction risk calculation device 210 connected by the network 313. Good.

  In the first to fourth embodiments, the case where 130 types of speech prosody features are used has been described as an example. However, the present invention is not limited to this, and other types of speech prosody features are used. Prosodic feature quantities may be used, and fewer or more types of speech prosodic feature quantities may be used.

  In addition, the case where the optimal speech prosodic feature value or combination of synthesis variables is searched using the forward stepwise method (variable increase / decrease method) has been described as an example. However, the present invention is not limited to this, and other sequential selections are possible. A method may be used to search for an optimal speech prosodic feature value or combination of synthetic variables. For example, an optimal speech prosodic feature value or combination of synthesis variables may be searched using a variable increase method, a variable decrease method, or the like. In addition, an optimal speech prosodic feature value or combination of synthetic variables may be searched using a simultaneous selection method such as an EM algorithm, a genetic algorithm (GA), or particle swarm optimization (PSO). Good.

Moreover, although the case where the weighting of each speech prosodic feature amount or each synthetic variable used in the learning model is learned by using multiple regression analysis has been described as an example, the present invention is not limited to this, for example, ridge regression Further, the weight of each speech prosodic feature quantity or each synthesis variable may be learned using a method such as support vector regression (SV regression), kernel regression, or the like.
In the above embodiment, the case where a numerical value corresponding to the score (0 to 30) of the cognitive function test such as the Hasegawa formula score is calculated as the risk level of cognitive dysfunction is described as an example. For example, the risk of cognitive dysfunction represented by three categories of normal (NL), suspected dementia (MCI), and dementia (AD) may be obtained. In this case, the risk of cognitive impairment is calculated using a learning model such as Bayesian network, canonical discriminant analysis, linear discriminant analysis, neural network, naive Bayes method, support vector machine (SVM), etc. That's fine. Further, the weight of each speech prosody feature amount or each synthesis variable used in the learning model may be determined by learning.

Moreover, although the case where the question is reproduced from the speaker has been described as an example, the present invention is not limited to this, and the question may be displayed by a display device.
Moreover, although the case where the risk level of cognitive dysfunction is calculated based on the voice data input as an answer to the question has been described as an example, the present invention is not limited to this, and the voice data input by voice monitoring is not limited thereto. Based on this, the risk level of cognitive dysfunction may be calculated.

10, 210, 312, 414 Cognitive impairment risk calculator 12 Voice input unit 14, 214, 314 Computer 20 Learning data storage unit 22 Feature amount selection unit 24, 224 Weight determination unit 25 Question playback unit 28 Feature amount extraction unit 30 , 230 Risk level calculation unit 221 Variable synthesis unit 222 Composite variable selection unit 229 Composite variable calculation units 310 and 410 Cognitive impairment disorder risk calculation system 311 Feature amount learning device

Claims (9)

Based on a plurality of types of prosodic feature quantities extracted from the voice data and a plurality of learning data including the risk level of cognitive impairment obtained for a speaker of the voice data, Feature quantity selection means for selecting a combination of the prosodic feature quantities that have the highest correlation with the risk,
Weight determination means for determining a weight for each of the selected combinations of prosodic feature quantities based on the combination of the selected prosodic feature quantities and the degree of risk of each of the plurality of learning data;
Feature quantity extraction means for extracting the plurality of types of prosodic feature quantities from the input voice data;
Based on the combination of the selected prosodic feature quantities of the prosodic feature quantities extracted by the feature quantity extracting means and the weights determined by the weight determining means, the risk level of cognitive dysfunction is calculated. Risk calculation means;
Cognitive impairment risk calculation device including Analysis processing on the plurality of types of prosodic feature quantities of the plurality of learning data including a plurality of types of prosodic feature quantities extracted from the speech data and a risk level of cognitive impairment obtained for a speaker of the speech data And a synthetic variable generating means for generating a plurality of synthetic variables obtained by synthesizing the plural types of prosodic feature quantities;
Based on the plurality of learning data and the plurality of types of generated synthetic variables, a synthetic variable selection that selects the combination of the synthetic variables that has the highest correlation with the degree of risk from the plurality of types of synthetic variables. Means,
Weighting determination for determining a weight for each of the selected combination of synthetic variables based on the combination of the synthetic variables obtained for each of the plurality of learning data and the risk of each of the plurality of learning data Means,
Feature quantity extraction means for extracting the plurality of types of prosodic feature quantities from the input voice data;
Risk calculation means for calculating the risk of cognitive impairment based on the combination of the synthetic variables obtained from the prosodic feature values extracted by the feature value extraction means and the weights determined by the weight determination means When,
Cognitive impairment risk calculation device including   The plurality of types of prosodic feature quantities include a feature quantity relating to a frequency component of speech, a feature quantity relating to a speech formant structure, a feature quantity relating to speech volume, a feature quantity relating to speech speed, and a feature relating to a response time until a question is answered The cognitive impairment risk calculating device according to claim 1 or 2, comprising at least one of the quantities.   The cognitive dysfunction risk degree calculation device according to any one of claims 1 to 3, wherein the risk level of the cognitive dysfunction of the learning data is determined by a Hasegawa simple intelligence evaluation scale for the speaker. .   The cognitive dysfunction risk degree calculation device according to any one of claims 1 to 4, wherein the feature amount extraction unit extracts the plurality of types of prosodic feature amounts from voice data input as an answer to a question. Based on a plurality of types of prosodic feature quantities extracted from the voice data and a plurality of learning data including the risk level of cognitive impairment obtained for a speaker of the voice data, Based on the feature quantity selection means for selecting the combination of the prosodic feature quantities that has the highest correlation with the risk degree, and the risk degree and the combination of the selected prosodic feature quantities of each of the plurality of learning data , A feature amount selection device including weight determination means for determining a weight for each of the selected prosodic feature amount combinations;
A feature amount extracting means for extracting the plurality of types of prosodic feature quantities from the input speech data; and a combination of the selected prosodic feature quantities among the prosodic feature quantities extracted by the feature quantity extracting means; A risk calculating device including a risk calculating means for calculating a risk of cognitive impairment based on the weight determined by the weight determining means;
Cognitive impairment risk calculation system. Analysis processing on the plurality of types of prosodic feature quantities of the plurality of learning data including a plurality of types of prosodic feature quantities extracted from the speech data and a risk level of cognitive impairment obtained for a speaker of the speech data And a synthetic variable generating means for generating a plurality of synthetic variables obtained by synthesizing the plural types of prosodic feature quantities,
Based on the plurality of learning data and the plurality of types of generated synthetic variables, a synthetic variable selection that selects the combination of the synthetic variables that has the highest correlation with the degree of risk from the plurality of types of synthetic variables. And a weight for each of the selected combination of synthesis variables is determined based on the combination of the combination variables obtained for each of the plurality of learning data and the risk of each of the plurality of learning data. A synthetic variable selection device including a weight determination means for
Feature quantity extraction means for extracting the plurality of types of prosodic feature quantities from the input speech data; a combination of the synthetic variables obtained from the prosodic feature quantities extracted by the feature quantity extraction means; and the weight determination means A risk level calculation device including a risk level calculation means for calculating a risk level of cognitive impairment based on the weight determined by
Cognitive impairment risk calculation system. Computer
Based on a plurality of types of prosodic feature quantities extracted from the voice data and a plurality of learning data including the risk level of cognitive impairment obtained for a speaker of the voice data, Feature quantity selection means for selecting a combination of the prosodic feature quantities that has the highest correlation with the risk level;
Weight determination means for determining a weight for each of the selected combination of prosodic feature quantities based on the combination of the selected prosodic feature quantities and the degree of risk of each of the plurality of learning data;
A feature amount extracting means for extracting the plurality of types of prosodic feature quantities from the input speech data; and a combination of the selected prosodic feature quantities among the prosodic feature quantities extracted by the feature quantity extracting means; A program for functioning as a risk degree calculating means for calculating a risk degree of cognitive dysfunction based on the weight determined by the weight determining means. Computer
Analysis processing on the plurality of types of prosodic feature quantities of the plurality of learning data including a plurality of types of prosodic feature quantities extracted from the speech data and a risk level of cognitive impairment obtained for a speaker of the speech data And a synthetic variable generating means for generating a plurality of synthetic variables obtained by synthesizing the plural types of prosodic feature quantities,
Based on the plurality of learning data and the plurality of types of generated synthetic variables, a synthetic variable selection that selects the combination of the synthetic variables that has the highest correlation with the degree of risk from the plurality of types of synthetic variables. means,
Weighting determination for determining a weight for each of the selected combination of synthetic variables based on the combination of the synthetic variables obtained for each of the plurality of learning data and the risk of each of the plurality of learning data means,
Feature quantity extraction means for extracting the plurality of types of prosodic feature quantities from the input speech data; a combination of the synthetic variables obtained from the prosodic feature quantities extracted by the feature quantity extraction means; and the weight determination means A program for functioning as a risk level calculation means for calculating the risk level of cognitive impairment based on the weight determined by.

Images