JP2017111760A - Emotion estimation device creation method, emotion estimation device creation device, emotion estimation method, emotion estimation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve estimation accuracy in estimation of an emotion of a speaker based on voice data.SOLUTION: An emotion estimation device creation device 100 comprises: an analysis zone setting part 120 for setting an analysis zone for analyzing a feature amount of voice data which is a source of teacher data; an accent type determination part 130 for determining a pattern in which the feature amount of the voice data included in the analysis zone is changed, as a change pattern of the feature amount of the voice data included in the analysis zone based on the change patterns classified into plural classes; and an emotion estimation device creation part 150 for creating an emotion estimation device for estimating an emotion of the speaker when the speaker produces speech, for every change pattern of the feature amount, with the voice data which is classified for every change pattern of the feature amount as the teacher data.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an emotion estimator generation method, an emotion estimator generation device, an emotion estimation method, an emotion estimation device, and a program.

  Development of a technique for estimating a speaker's emotion using an emotion estimation device generated by machine learning using a voice data group in which emotions are labeled as teacher data is underway. For example, Patent Literature 1 discloses a technique for obtaining respective amounts of change in speech intensity, speech tempo, and speech inflection, and estimating a speaker's emotion based on the obtained change amount.

JP 2002-91482 A

  In general, speaking in an excited state tends to be faster and louder than normal speech. In addition, when speaking in a discouraged state, speaking tends to be slower than normal speech, and the voice tends to be low. Thus, there is a correlation between the emotion of the speaker at the time of speaking and the feature amount of the voice. Patent Document 1 discloses a technique for estimating the emotion of a speaker by analyzing such a change in the feature amount of voice data.

  By the way, when comparing the feature quantity of speech uttered in a normal emotional state with the feature quantity of speech uttered in an angry emotional state, the tendency of changes in feature quantities may differ between short words and long words. is there. For example, a short word that is easy to utter often has a large change in voice feature value depending on the emotional state at the time of utterance. On the other hand, a long word that is difficult to utter, such as a quick-speaking word, may have a small change in voice feature value depending on the emotional state at the time of utterance. Since the technique disclosed in Patent Document 1 estimates a speaker's emotion uniformly by using words that have a small change in voice feature amount and words that have a large change depending on the emotional state at the time of utterance, the estimation accuracy increases. There was a problem that it was difficult.

  The present invention has been made in view of such circumstances, and an emotion estimator generation method, an emotion estimator generation device, and an emotion estimation method capable of improving the estimation accuracy for estimating the emotion of a speaker from speech data An object is to provide an emotion estimation apparatus and program.

In order to achieve the above object, an emotion estimator generation method according to the first aspect of the present invention includes:
An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the voice data that is the source of the teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Steps,
It is characterized by including.

Moreover, the emotion estimator generation device according to the second aspect of the present invention provides:
An analysis interval setting means for setting an analysis interval for analyzing a feature amount of voice data that is a source of teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Means,
It is characterized by including.

A program according to the third aspect of the present invention is:
Analysis interval setting means for setting an analysis interval for analyzing features of speech data that is a source of teacher data on a computer;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data means,
It is made to function as.

Moreover, the emotion estimation method according to the fourth aspect of the present invention includes:
An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation step;
It is characterized by including.

An emotion estimation apparatus according to the fifth aspect of the present invention is:
Analysis interval setting means for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation means;
It is characterized by providing.

  ADVANTAGE OF THE INVENTION According to this invention, the estimation precision which estimates a speaker's emotion from audio | voice data can be improved.

It is a block diagram which shows the physical structure of the emotion estimator production | generation apparatus which concerns on Embodiment 1 of this invention. It is a block diagram which shows the function structure of the emotion estimator production | generation apparatus which concerns on Embodiment 1 of this invention. It is a figure for demonstrating a morpheme. It is a figure for demonstrating a mora area. It is a figure for demonstrating the analysis method of a feature-value. It is a figure for demonstrating the analysis method of a feature-value. It is a figure for demonstrating classification. It is a figure for demonstrating the image of the identification threshold value of the produced | generated emotion estimation apparatus. It is a block diagram which shows the function structure of an emotion estimation apparatus. It is a flowchart for demonstrating the production | generation process of an emotion estimator. It is a flowchart for demonstrating an emotion estimation process. It is a figure for demonstrating the analysis area of the feature-value which concerns on the modification 1. FIG. It is a figure for demonstrating the analysis method of the feature-value which concerns on the modification 1. FIG. FIG. 10 is a diagram for explaining analysis of a feature amount based on sound intensity according to Modification Example 2. It is a figure for demonstrating the technique which estimates the degree of the some emotion which concerns on the modification 5. FIG.

  Hereinafter, an emotion estimator generation method, an emotion estimator generation device, an emotion estimation method, an emotion estimation device, and a program according to an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals.

(Embodiment 1)
In the present embodiment, after describing an emotion estimator generating device that generates an emotion estimator that estimates an emotion of a speaker from speech data, an emotion estimating device that estimates the emotion of the speaker when speaking the speech will be described. . In this embodiment, the emotion estimation device is in a state of sadness (sadness), bored state (boring), angry state (anger), surprised state (surprise), and discouraged. A case where it is estimated that the state is one of the seven basic emotional states, that is, a state of being disappointed (disappointed), a state of being disgusted (disgust), and a state of being happy (joy).
In the following embodiment, the change pattern of the feature amount of the audio data is referred to as an accent type.

  The emotion estimator generation device 100 according to the first embodiment physically includes a control unit 1, a storage unit 2, an input / output unit 3, and a bus 4 as illustrated in FIG. 1.

  The control unit 1 includes a ROM (Read Only Memory), a RAM (Random Access Memory), and a CPU (Central Processing Unit). The ROM stores an emotion estimator generation program according to the present embodiment, an initial program for performing various initial settings, hardware inspection, program loading, and the like. The RAM functions as a work area for temporarily storing various software programs executed by the CPU and data necessary for executing these software programs. The CPU is a central processing unit that executes various processes and operations by executing various software programs.

  The storage unit 2 includes a nonvolatile memory such as a hard disk drive or a flash memory. The storage unit 2 stores voice data and the like as teacher data.

  The input / output unit 3 includes an audio input device for acquiring audio data as teacher data, a CD (Compact Disc) drive, and a USB (Universal Serial Bus) interface. The input / output unit 3 acquires audio data as teacher data. The input / output unit 3 outputs the generated emotion estimator to a program or a parameter that determines the characteristics of the emotion estimator.

  The bus 4 connects the control unit 1, the storage unit 2, and the input / output unit 3.

  As shown in FIG. 2, the emotion estimator generation device 100 functionally includes a voice data acquisition unit 110, an analysis interval setting unit 120, an accent type determination unit 130, a feature amount extraction unit 140, and emotion estimation. Generator generator 150. The analysis section setting unit 120 includes a morpheme analysis unit 121 and an accent phrase extraction unit 122. Further, the accent type determination unit 130 includes a mora section extraction unit 131 and an accent type extraction unit 132.

  The voice data acquisition unit 110 acquires voice data used as teacher data in order to generate an emotion estimator via the input / output unit 3. The teacher data is composed of speech uttered in seven emotional states, for example, sadness, boredom, anger, surprise, discouragement, disgust, and joy. The teacher data is composed of sound data including a sufficiently large number of types of words. It is preferable that the number of utterances of teacher data and the types of accent phrases included in the teacher data are large. An accent phrase is a unit that classifies speech data in which a noun and a particle or a verb and an auxiliary verb are combined. For example, as teacher data, voice data including more than 1000 types of accent phrases that a large number of people of about 500 uttered in 7 types of emotion states is prepared.

  The analysis section setting unit 120 sets an analysis section which is a unit for analyzing the characteristics of the voice data used as teacher data. For this purpose, the analysis section setting unit 120 includes a morpheme analysis unit 121 and an accent phrase extraction unit 122.

  The morpheme analyzer 121 divides the acquired voice data into morphemes. A morpheme is a minimum unit having a meaning as a language. For example, as shown in FIG. 3, the voice that “the shaved drawing of the shaved figure well” is shown in FIG. 3 as “shaved”, “ga”, “folding screen”, “ni”, “skilled”, “ni”. , “Shaved”, “No”, “Picture”, “O”, “Draw”, “Ta”.

  The accent phrase extraction unit 122 extracts an accent phrase from the acquired voice data. An accent phrase is a section in which a noun or verb divided into morphemes is combined with a subsequent particle or auxiliary verb. In the above-mentioned example, the accent phrases are “shaved is”, “is in a folding screen”, “is well”, “is a shaved”, “draws”, “draws”. In the present embodiment, a case will be described in which the characteristics of audio data are analyzed in units of accent phrases. The reason why voice data is analyzed in units of accent phrases is that the emotional state of the speaker often changes in units of accent phrases.

  The accent type determination unit 130 determines the accent type of the accent phrase. The accent type is “H” when the voice feature amount is larger than the average feature amount and “L” when the speech feature amount is smaller than the average feature amount for each mora section in which the syllable constituting the accent phrase is spoken. Is a combination pattern of “H” and “L” obtained. In order to determine the accent type, the accent type determination unit 130 includes a mora section extraction unit 131 and an accent type extraction unit 132.

  As illustrated in FIG. 4, the mora section extraction unit 131 extracts a mora section from the speech data of the accent phrase section to be analyzed. A mora section is a section in which one syllable is spoken. In the case of the accent phrase “shaved”, each syllable of “BO”, “U”, “Z”, and “GA” is a mora section.

  The accent type extraction unit 132 determines whether each of the mora sections corresponds to “H” or “L”, and extracts the accent type of the accent phrase. As an accent type extraction method, there is a method of using feature quantities such as voice intensity, voice pitch, and phoneme speech duration. Here, an extraction method that focuses on the pitch of speech will be described with reference to FIGS.

  As shown in FIG. 5, the accent type extraction unit 132 sets a time window of a predetermined time for further subdividing the mora section. A window 1 is set for one mora section, and audio data in the window is subjected to FFT (Fast Fourier Transform) conversion. Next, the audio data in the window 2 in which the window 1 is shifted by a predetermined time dt is subjected to FFT conversion. Thereafter, the audio data in the window n is similarly subjected to FFT conversion. As the time window setting method, for example, a time width dt to be shifted from the time window is set so that ten or more time windows are formed in the mora section. This is because if the number of time windows is too small, the calculation accuracy decreases.

  FIG. 6 is an example showing the spectral distribution of the audio data in each window obtained by the above-described FFT conversion. The horizontal axis is frequency, and the vertical axis is spectrum intensity. The peak frequency existing in the lowest frequency region in this spectrum is defined as f0. This f0 indicates the fundamental frequency unique to the speaker obtained from the audio data of the window section. Let f0 obtained from window 1 be f0_1, and f0 obtained from window 2 be f0_2. Similarly, let f0 obtained from the window n be f0_n. And the accent type | mold extraction part 132 calculates the average value from f0_1 to f0_n, and makes it the average fundamental frequency 1_f0 of a 1st mora area.

  The accent type extraction unit 132 performs the same calculation for all the mora sections included in the accent phrase. The average fundamental frequency m_f0 of the m-th mora section can be calculated using Equation 1.

      m_f0 = 1 / n ・ Σf0_n (Formula 1)

  Next, the accent type extraction unit 132 obtains the average fundamental frequency m_th in the accent phrase section using Equation 2.

      m_th = {max (1_f0, ..., n_f0) -min (1_f0, ..., n_f0)} / 2 (Formula 2)

  Next, the accent type extraction unit 132 compares the average fundamental frequency m_f0 of the mora section with the average fundamental frequency m_th of the accent phrase section, and is “H” if m_f0 ≧ m_th, and “L” if m_f0 <m_th. Is assigned to each mora section. The accent type extraction unit 132 extracts an accent type composed of a combination of H and L by assigning “H” and “L” to each mora section constituting the accent phrase in this way.

  The accent type determination unit 130 performs this process for all accent phrases generated from the teacher data. The example of the accent type shown in FIG. 7 is an example in which 20 types of accent types are selected in descending order of occurrence frequency among the accent types obtained from a large amount of teacher data, and the number of mora sections is 6 or less. When the number of mora sections included in the accent phrase to be analyzed is 6 or less, class names from class 1 to class 20 are assigned to the 20 types of accent types in order. There is a one-to-one correspondence between accent types and classes. The accent type determination unit 130 stores the accent phrase unit speech data as the teacher data and the accent type (class) in the storage unit 2 in association with each other. Note that accent phrases that do not correspond to these 20 types of accent types are excluded from the teacher data. Hereinafter, in the present embodiment, a case where the number of mora sections included in an accent phrase to be analyzed is 6 or less will be described.

  Here, when the number of mora sections included in an accent phrase is 6 or less, 20 types of accent types are selected based on the result of statistical processing of a large amount of Japanese spoken with 7 types of emotions. Twenty accent types are selected in descending order of frequency. If the number of classes of the accent type is reduced, the frequency of teacher data that does not correspond to the accent type increases, and the estimation accuracy of the emotion estimation device incorporating the generated emotion estimator decreases. On the other hand, when the number of accent-type classes is increased, the number of types of emotion estimators to be generated increases, and the manufacturing cost of the emotion estimation device increases. Therefore, the number of accent-type classes is determined based on the balance between the two. Note that the 20 accent types are examples in the case of Japanese, where the number of mora sections included in the accent phrase is 6 or less. When the number of mora sections included in the accent phrase is 7 or more, or in other languages, the frequency of occurrence of the accent type needs to be further statistically determined.

  Returning to FIG. 2, the feature amount extraction unit 140 extracts a feature amount of speech for each accent phrase. The feature amount of speech is the size of speech, the pitch of speech, the utterance time length of phonemes, and the like. Then, class names of class 1 to class 20 determined by the accent type determination unit 130 are given to the extracted feature amount, and stored in the storage unit 2 as teacher data.

  The emotion estimator generation unit 150 acquires teacher data from the storage unit 2 for each class, and generates an emotion estimator adapted to each class. Specifically, the emotion estimator generation unit 150 acquires teacher data classified into class 1, and the emotion state at the time of utterance of the teacher data is a sad state (sadness) that is seven types of emotion states. , Bored (bored), angry (angry), surprised (surprise), discouraged (disappointed), disgusted (disgust), happy A class 1 emotion estimator that classifies (joy) is generated. FIG. 8 is an image diagram in which the identification threshold value for identifying the seven types of emotions is expressed in two dimensions. A known technique can be used as a method for generating an emotion estimator based on teacher data. Next, the emotion estimator generation unit 150 acquires the teacher data classified into class 2, and changes the emotion state at the time of utterance of the teacher data into seven kinds of emotion states, sadness, boredom, anger, surprise, discouragement. Generate an emotion estimator for class 2 that classifies as hate, joy. Similarly, the emotion estimator generation unit 150 generates emotion estimators up to class 20.

  Next, the configuration of the emotion estimation apparatus 200 incorporating the 20 types of emotion estimators generated by the above description will be described with reference to FIG. The emotion estimation apparatus 200 is physically the same as the configuration shown in FIG.

  The ROM included in the control unit 1 stores an emotion estimation program generated by the emotion estimator generation device 100 according to the present embodiment. The storage unit 2 stores audio data and the like to be analyzed. The input / output unit 3 includes an audio input device for acquiring audio data to be analyzed, a CD drive, and a USB interface. Further, the input / output unit 3 may acquire a parameter that determines the characteristics of the emotion estimator generated by the emotion estimator generation device 100. Further, the input / output unit 3 includes a display device or a sound output device for outputting the result of estimating the emotion.

  As shown in FIG. 9, the emotion estimation apparatus 200 includes a voice data acquisition unit 210, a speaker division unit 220, an analysis section setting unit 230, an accent type determination unit 240, a selection unit 250, and a feature amount extraction unit. 260, the emotion estimation unit 270, and the integration unit 280. The analysis section setting unit 230 includes functions of a morphological analysis unit 231 and an accent phrase extraction unit 232. Further, the accent type determination unit 240 includes functions of a mora section extraction unit 241 and an accent type extraction unit 242.

  The voice data acquisition unit 210 acquires voice to be analyzed which is spoken by the user. The audio data acquisition unit 210 includes an audio acquisition device such as a microphone. The audio data acquisition unit 210 includes a CD drive and a USB interface, and can acquire the user's audio as audio data.

  The speaker dividing unit 220 divides the acquired voice data to be analyzed for each speaker. This is to estimate the speaker's emotion for each sentence spoken by a single speaker when the voice data includes a plurality of voice data. A method of dividing the voice data for each speaker is performed using a known technique. For example, the division can be performed based on the correlation such as the intensity of the voice, the pitch of the voice, and the utterance time length of the phoneme.

  The analysis section setting unit 230 and the accent type determination unit 240 have the same configuration as the emotion estimator generation device 100 in order to analyze the speech data to be analyzed under the same conditions as when the emotion estimator is generated. That is, the analysis section setting unit 230 extracts an accent phrase from the audio data, and the accent type determination unit 240 determines an accent type (class) for each accent phrase.

  The selection unit 250 selects an emotion estimator corresponding to the corresponding class for each classified accent phrase. Specifically, the emotion estimator corresponding to the class of the accent phrase to be analyzed is selected from the 20 types of emotion estimators built in the emotion estimation device 200.

  The feature amount extraction unit 260 has the same configuration as the emotion estimator generation device 100, and extracts feature amounts from the speech data under the same conditions. Then, the feature quantity extraction unit 260 stores the extracted feature quantity and the class name indicating the accent type in association with each other in the storage unit 2.

  The emotion estimation unit 270 uses the emotion estimator selected by the selection unit 250 to estimate the speaker's emotion for each accent phrase. Specifically, when estimating emotions of accent phrases classified into class 1, emotion estimation unit 270 selects an emotion estimator for class 1 and estimates the emotion of the speaker. When estimating emotions of accent phrases classified into class n, emotion estimation unit 270 selects an emotion estimator for class n and estimates the emotion of the speaker. Then, the emotion estimation unit 270 estimates whether the speaker's emotional state when the accent phrase to be analyzed corresponds to one of sadness, boredom, anger, surprise, discouragement, disgust, and joy. To do.

  The integration unit 280 estimates the emotion of the speaker for each sentence of the audio data. Specifically, the integration unit 280 estimates the emotion that is most common in one sentence as the emotion of the speaker who uttered the sentence. For example, “Bow is a good screen” composed of seven accent phrases: “Bow is”, “Beautiful”, “Good”, “Shaved”, “Draw”, “Draw”, “Ta” In the sentence “Drawn a shaved picture,” the number of accent phrases identified as “joy” was 4, the number of accent phrases identified as “anger” was 2, and it was determined as “surprise”. If the number of accent phrases is 1, the most “joy” is estimated as the emotion of the speaker when this “shaven skillfully painted the shaved picture”.

  Next, a process in which the emotion estimator generation device 100 having the above configuration generates an emotion estimator will be described with reference to FIG. It is assumed that voice data uttered in seven kinds of emotional states such as sadness, boredom, anger, surprise, discouragement, disgust, and joy used as teacher data is stored in the storage unit 2 in advance. It is assumed that the number of mora sections of the accent phrase included in the speech data to be analyzed is 6 or less. The person in charge of generating the emotion estimator starts the emotion estimator generation program installed in advance in the emotion estimator generation device 100, whereby the flowchart shown in FIG. 10 is started.

  When the emotion estimator generation program is activated, the control unit 1 acquires the teacher data stored in the storage unit 2 in the voice data acquisition unit 110 (step S11). Then, the morpheme analysis unit 121 divides the acquired voice data in units of morphemes (step S12). Next, the accent phrase extraction unit 122 extracts an accent phrase, which is a unit for analyzing the characteristics of the voice data, and divides the voice data into accent phrases (step S13).

  Next, the mora section extraction unit 131 extracts a mora section included in the accent phrase (step S14). Then, the accent type extraction unit 132 extracts the accent type of the accent phrase (step S15). Specifically, as described with reference to FIG. 7, the accent type extraction unit 132 classifies accent phrases used as teacher data into 20 classes (step S16). As described with reference to FIGS. 5 and 6, the accent type extraction unit 132 compares the average basic frequency m_f0 for each mora section with the average basic frequency m_th of the accent phrase section, as described with reference to FIGS. 5 and 6. If ≧ m_th, “H” is assigned to each mora section, and if m_f0 <m_th, “L” is assigned to each mora section. The accent type extraction unit 132 assigns H and L to the accent phrase used as teacher data in this way for each mora section constituting the accent phrase, and extracts the accent type based on the H and L patterns. Then, the accent type determination unit 130 determines the accent type of the teacher data as one of the 20 accent types (classes) shown in FIG.

  Next, the feature amount extraction unit 140 extracts the feature amount of the speech data as the teacher data for each accent phrase, and associates the extracted feature amount data with the classified class in the storage unit 2 as teacher data. Store (step S17).

  The emotion estimator generation unit 150 generates an emotion estimator for each class based on the teacher data classified for each accent type (class) (step S18). Specifically, the emotion estimator generation unit 150 acquires teacher data (accent phrases) classified into class 1 and sets the emotional state at the time of utterance of the teacher data as sadness as seven types of emotional states, Generate emotion estimators that can be categorized as boredom, anger, surprise, discouragement, disgust, and joy. More specifically, identification threshold values (parameters of mathematical expressions constituting the classifier) for classification into seven types of emotions as shown in FIG. 8 are generated. Next, the emotion estimator generation unit 150 acquires the teacher data (accent phrase) classified into class 2, and changes the emotional state at the time of utterance of the teacher data to sadness, boredom, A second emotion estimator that can be classified into anger, surprise, discouragement, disgust, and joy is generated. The emotion estimator generation unit 150 generates 20 types of emotion estimators in this way. Above, description of the emotion estimator production | generation process of the emotion estimator production | generation apparatus 100 is complete | finished.

  Next, an emotion estimation process in which the emotion estimation device 200 including the 20 types of emotion estimators generated by the emotion estimator generation device 100 estimates the emotion of the speaker will be described with reference to FIG. The flowchart shown in FIG. 11 is started when the user starts an emotion estimation program installed in advance in emotion estimation apparatus 200 and inputs voice data to be analyzed to emotion estimation apparatus 200.

  When the emotion estimation program is activated and the user inputs voice data to be analyzed to the emotion estimation apparatus 200, the controller 1 divides the acquired voice data for each speaker and stores the data in the storage unit 2. To remember. Next, the analysis section setting unit 230 acquires voice data for each speaker from the storage unit 2 (step S31). Next, the morphological analysis unit 231 decomposes speech data of an arbitrary speaker into morphemes (step S32), and the accent phrase extraction unit 232 determines an accent phrase that is an analysis unit (step S33).

  Next, the accent type determination unit 240 determines an accent type (class) for each accent phrase as in the operation description of the emotion estimator generation device 100. Specifically, the mora section extraction unit 241 extracts the mora section included in the corresponding accent phrase (step S34), and the accent type extraction unit 242 extracts the accent type of the accent phrase (step S35). Then, the accent type determination unit 240 determines the class to which the accent phrase belongs from the extracted accent types. And the selection part 250 is an emotion estimator production | generation apparatus 100 based on the teacher data of the same accent type (class) as an emotion estimator used in order to estimate the speaker's emotion when uttering an applicable accent phrase. The emotion estimator generated by is selected (step S36).

  On the other hand, the feature quantity extraction unit 260 extracts voice feature quantities such as voice intensity, voice pitch, and phoneme duration length of the accent phrase to be analyzed, and associates the extracted feature quantities with the determined class. It memorize | stores in the memory | storage part 2 (step S37).

  Next, the emotion estimation unit 270 uses the emotion estimator selected by the selection unit 250 to estimate the emotion of the speaker when speaking the corresponding accent phrase (step S38).

  When emotion estimation is completed for one of the accent phrases, emotion estimation apparatus 200 determines whether there is an accent phrase that has not been analyzed (step S39). If there is an accent phrase that has not been analyzed (step S39: No), another accent phrase that has not been analyzed is extracted (step S40), and an emotion corresponding to the accent phrase is estimated.

  If the analysis of all accent phrases has been completed (step S39: Yes), the emotion estimation apparatus 200 performs integration processing for each analyzed sentence (step S41). Specifically, based on the emotion estimation result for each accent phrase included in the sentence to be analyzed, the integration unit 280 estimates the most common emotion as the speaker's emotion when the sentence is uttered.

  Next, the emotion estimation apparatus 200 determines whether or not the analysis has been completed for all sentences spoken by the first acquired arbitrary person (step S42). If the analysis has not been completed for all sentences (step S42: No), other sentences are extracted (step S43), and the emotion estimation process is continued for the other sentences.

  On the other hand, if the analysis has been completed for all sentences (step S42: Yes), emotion estimation apparatus 200 determines whether emotion estimation has been completed for all persons included in the voice data. (Step S44). If the analysis has not been completed for all the people (step S44: No), the voice estimation data of other people is extracted and the emotion estimation process is continued (step S45). If the analysis process has been completed for all persons (step S44: Yes), the emotion estimation process is terminated.

  As described above, the emotion estimator generation device 100 generates an emotion estimator for each accent type based on the teacher data classified for each accent type. And the emotion estimation apparatus 200 estimates the speaker's emotion using the emotion estimator generated for each accent type. Specifically, emotion estimation apparatus 200 classifies speech data to be analyzed for each accent type, and estimates a speaker's emotion using an emotion estimator generated based on teacher data having the same accent type. . Thereby, the estimation precision which estimates a speaker's emotion from audio | voice data can be improved.

  In addition, since the accent type extraction unit 132 extracts changes in the feature amount of speech in units of mora sections, the emotion estimator generation device 100 uses an emotion estimator that can analyze the speaker's emotion in more detail. Can be generated.

  The accent type extraction unit 132 extracts an accent type based on a change in the fundamental frequency of the voice. The fundamental frequency of speech tends to change depending on the emotional state during speech. Therefore, the emotion estimator generation device 100 can generate an emotion estimator that can estimate the speaker's emotion more accurately. For the same reason, the emotion estimation apparatus 200 can estimate the speaker's emotion more accurately.

  Since the analysis section setting unit 120 analyzes the speech in units of morphemes, the emotion estimator generation device 100 can generate an emotion estimator that can analyze the speaker's emotion more accurately.

  The emotion estimator generation device 100 generates an emotion estimator that classifies emotions of a speaker when speaking into sadness, boredom, anger, surprise, discouragement, disgust, and joy. The emotion estimation apparatus 200 incorporating this estimator can classify the emotional state of the speaker when speaking into sadness, boredom, anger, surprise, discouragement, disgust, and joy.

(Modification 1)
In the first embodiment, the feature amount is analyzed in units of mora sections in order to determine the accent type. In the first modification, a description will be given of analyzing the feature quantity limited to the vowel section in the mora section. Specifically, as shown in FIG. 12, the voice data of only the vowel section is taken out and the feature amount is analyzed as shown in FIG. The fundamental frequency analysis method is the same as that described in the first embodiment.

  The reason for focusing only on the vowel interval is that the vowel duration is longer than the consonant interval, and the energy of the included speech is larger than the consonant interval. This is because it appears prominently in the section.

  As described above, the emotion estimation apparatus 200 according to the modification 1 can improve the estimation accuracy of emotion estimation by analyzing the feature amount only in the vowel section.

(Modification 2)
In the description of the first embodiment, the case where the accent type extraction unit 132 uses the pitch information (sound fundamental frequency) as the feature amount of the sound has been described. In the second modification, a case will be described in which voice intensity information is used as a voice feature amount. Here, a technique that focuses on the fact that the energy distribution of the voice in the vowel utterance section changes depending on the emotional state during utterance will be described.

  Specifically, the accent type extraction unit 132 determines whether the peak of the voice energy indicated by the dotted circle in FIG. 14 is in the first half or the second half of the vowel section. For example, “H” is assigned when a peak exists in the first half, and “L” is assigned when a peak exists in the second half. Thereby, the accent type determination unit 130 determines an accent type. When classifying accent types according to the strength of speech, it is necessary to examine the classification method based on experimental data. Other explanations are the same as those of the first embodiment.

  In the description of the second modification, the analysis method focusing on the change in the time position of the peak point of the voice energy has been described, but the accent type can be extracted using the change in the intensity of the voice. When speaking in an angry state, the strength of the voice tends to increase, and when speaking in a sad state, the strength of the voice tends to decrease, so this tendency is used. In this case, for example, the peak intensity of the voice for each mora section included in the accent phrase section is measured to obtain the average peak intensity of the accent phrase section. Then, the accent type can be extracted by comparing the peak intensity and the average peak intensity of the sound in the mora section and assigning “H” or “L” to each mora section.

  As described above, the emotion estimation apparatus 200 according to the modified example 2 analyzes the emotion state at the time of speech utterance using the change information of the strength of the speech, so that the estimation accuracy of emotion estimation can be improved.

(Modification 3)
In the third modification, a case will be described in which the phoneme duration is used as a voice feature. When speaking in an angry or happy state, the phoneme duration tends to be short, and when speaking in a boring or sad state, the phoneme duration tends to be long. is there.

  Specifically, the accent type extraction unit 132 compares the duration of the vowels included in the mora section with the average duration of the same vowel included in the teacher data, and the duration of the vowels included in the mora section If the length is longer than the average duration time, “H” is assigned, and if the length is shorter, “L” is assigned, whereby the accent type determination unit 130 determines the accent type.

  In the description of the first embodiment, the first modified example, and the second modified example, the average value of the voice feature amount in the accent phrase section, which is the analysis section, is compared with the average value in the mora section. However, when comparing with the phoneme duration length, the accent type extraction unit 132 of the emotion estimator generation device 100 does not calculate the average duration length of the teacher data but the average duration length of the entire teacher data instead of the average of the speech data in the analysis section. Compare with Since the duration is different depending on the vowel, it cannot be compared with the duration of different vowels. Since the number of the same vowels included in the accent phrase is small, the variation in the average duration time becomes large and causes erroneous determination. Therefore, it is preferable to average the entire teacher data.

  On the other hand, it is preferable that the accent type extraction unit 242 of the emotion estimation apparatus 200 calculates the average duration for each vowel for the voice data for each speaker classified by the speaker classification unit 220.

  When classifying accent types according to phoneme durations, it is necessary to consider how to classify them based on experimental data. Other explanations are the same as those of the first embodiment.

  As described above, the emotion estimation apparatus 200 according to the modified example 3 analyzes the emotional state at the time of speech utterance using the change information of the phoneme utterance time length, so that the estimation accuracy of emotion estimation can be improved. .

(Modification 4)
In the first embodiment and the first modification, the technique for extracting the accent type using the pitch information of the voice as the voice feature amount has been described. In the second modification, a technique for extracting an accent type using the intensity information of speech was introduced. In the third modification, a technique for extracting an accent type using the duration of phonemes was introduced. When extracting an accent type, these techniques can be used alone, but an accent type can also be extracted by combining two or more techniques such as voice pitch information and voice intensity information. Combining two or more pieces of information increases the number of accent types, but can improve the accuracy of emotion estimation.

  In the above description, the sound intensity, the pitch of the sound, and the duration of the phoneme have been described as examples of the feature amount of the sound. However, the present invention is not limited to this. For example, the accent type can be determined by extracting the change amount of the sound intensity, the change amount of the sound pitch, the change amount of the phoneme duration, and the like.

(Modification 5)
In the description of the first embodiment, a technique has been described in which, based on the emotion estimation result for each accent phrase included in the sentence to be analyzed, the most common emotion is estimated as the speaker's emotion when the sentence is spoken. . However, the method of integration processing need not be limited to this. For example, estimation including a plurality of emotions such as “joy with a little surprise” can be performed. In a classifier that constitutes an emotion estimator, it is often the case that a feature quantity is acquired as a vector and the emotion to be classified is determined based on the distance between the vector and an identification threshold. For example, the feature amounts corresponding to seven accent phrases of “shaved is”, “bold in the screen”, “skilled”, “shaved”, “draw”, and “drawn” are represented by 1 to 7 shown in FIG. It is assumed that the average vector obtained by combining the seven position vectors is the position vector indicated by “average” in FIG. In this case, the position vector “average” belongs to the joy region, but exists near the boundary between joy and surprise. In such a case, an emotion estimation result including a nuance that “a little surprising emotion may be mixed” may be output.

  Since FIGS. 8 and 15 are two-dimensional image representations of a 7-dimensional identification space, it is difficult to represent a complex example. However, in the classifier constituting the emotion estimator, it is possible to calculate the distance from each identification boundary numerically. Accordingly, it is possible to numerically calculate a combination of a plurality of emotions and a degree of the emotion (distance from the identification boundary) such as “anger and sadness” and “anger and discouragement”. Furthermore, by setting a plurality of thresholds, emotion estimation including two or more emotions such as “anger, sadness, and discouragement” is possible. Moreover, the composite degree of several emotions can also be estimated.

  By providing the configuration and processing described in the modification example 5, the emotion estimator generation device 100 can generate an emotion estimator that can estimate the degree of a plurality of emotions. Also, the emotion estimation device 200 including an emotion estimator capable of estimating a plurality of emotion levels can estimate a plurality of emotion levels of a speaker.

  In the description of the first embodiment, the emotional state of the speaker is classified into seven types of sadness, boredom, anger, surprise, discouragement, disgust, and joy. However, the emotion classification method needs to be limited to this. There is no. For example, it may be classified into four types of joy, anger, sorrow, and comfort.

  In the description of the first embodiment, the emotional state at the time of speaking of the speaker is described as being classified into any of the seven emotional states of sadness, boredom, anger, surprise, discouragement, disgust, and joy. Explained that teacher data that does not fall under is excluded. However, the emotional state of “normal” may be provided as the emotion of the speaker, and the teacher data that could not be classified into the seven emotions may be classified as the emotion “normal”. As a result, the emotion estimator generation device 100 can generate an emotion estimator that can estimate the emotion of the speaker to eight emotions obtained by adding “normal” to the seven emotions. The emotion estimation apparatus 200 can estimate the speaker's emotions as eight emotions by adding “normal” to the seven emotions.

  In the description of the first embodiment, the analysis interval is described as the unit of the accent phrase. However, the analysis interval is not necessarily limited to this. For example, the analysis interval may be a word utterance interval, a breathing interval interval that is a breathing interval, or a sentence utterance interval. When the analysis section is a sentence utterance section, the integration unit 280 may estimate a speaker's emotion in units of sentences, or may further estimate a speaker's emotion in units of a plurality of sentences. It may be.

  In the description of Expression 1, processing that represents the feature amount of the section using the average value has been described. However, processing may be performed using the median instead of the average value. Further, the processing may be performed using the lowest frequency as a representative value.

  In the description of Expression 2, processing that represents the feature amount of the section using the median value has been described, but processing may be performed using an average value instead of the median value.

  In addition to being able to provide the emotion estimator generation device 100 and the emotion estimation device 200 that have a configuration for realizing the functions according to the present invention in advance, by applying a program, existing personal computers, information terminal devices, and the like can be provided. It is also possible to function as the emotion estimator generation device 100 and the emotion estimation device 200 according to the present invention. That is, a program for realizing each functional configuration by the emotion estimator generation device 100 and the emotion estimation device 200 exemplified in the above embodiment can be executed by a CPU or the like that controls an existing personal computer or information terminal device. By applying, it can function as the emotion estimator generation device 100 and the emotion estimation device 200 according to the present invention. Further, the emotion estimator generation method and the emotion estimation method according to the present invention can be implemented using the emotion estimator generation device 100 and the emotion estimation device 200.

  Moreover, the application method of such a program is arbitrary. For example, the program can be stored and applied to a computer-readable recording medium (CD-ROM (Compact Disc Read-Only Memory), DVD (Digital Versatile Disc), MO (Magneto Optical disc), etc.), the Internet, etc. It is also possible to apply the program by storing it in a storage on the network and downloading it.

  As mentioned above, although preferable embodiment of this invention was described, this invention is not limited to the specific embodiment which concerns, This invention includes the invention described in the claim, and its equivalent range It is. Hereinafter, the invention described in the scope of claims of the present application will be appended.

(Appendix 1)
An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the voice data that is the source of the teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Steps,
An emotion estimator generation method including:

(Appendix 2)
The change pattern determining step includes:
A mora section extraction step for dividing the speech data included in the analysis section into mora sections that are syllable units;
The speech data of the analysis section is uttered based on the comparison result of comparing the average value of the feature amount of the speech data in the analysis section and the average value of the feature amount of the speech data in the mora section for each mora section. A change pattern extraction step for extracting a change pattern of the feature amount of the sound that changes for each mora section;
The emotion estimator generation method according to appendix 1, characterized by comprising:

(Appendix 3)
In the change pattern extraction step, the fundamental frequency of speech extracted from speech data is used as the feature amount of speech, and the average fundamental frequency of speech in the analysis interval and the average fundamental frequency of speech in the mora interval are determined for each mora interval. Compared to the above, when the average fundamental frequency of the voice in the mora section is higher than the average fundamental frequency of the voice in the analysis section, High is given, and when it is low, Low is given, and the voice changes to High and Low for each mora section. Extract the change pattern of the feature amount of
The method for generating an emotion estimator according to Supplementary Note 2, wherein:

(Appendix 4)
In the change pattern extraction step, the voice intensity extracted from the voice data is used as the voice feature value, and the average voice intensity in the analysis section and the average voice intensity in the mora section are compared for each mora section. When the average intensity of the voice in the mora section is higher than the average intensity of the voice in the analysis section, High is given, and when the average intensity is low, Low is given, and the change in the feature amount of the voice that changes between High and Low for each mora section Extract patterns,
The method for generating an emotion estimator according to Supplementary Note 2, wherein:

(Appendix 5)
In the change pattern extraction step, the phoneme duration extracted from the speech data is used as the feature amount of speech, the average duration of phonemes in the analysis interval, and the average duration of phonemes in the mora interval, Compared for each mora section, High is given when the average duration of phonemes in the mora section is longer than the average duration of phonemes in the analysis section, Low is given when the phoneme is short, and High for each mora section. Extracting the change pattern of the feature amount of the voice that changes to Low,
The method for generating an emotion estimator according to Supplementary Note 2, wherein:

(Appendix 6)
In the change pattern extraction step, a change pattern of the speech feature value is extracted using at least one of the fundamental frequency of the speech, the strength of the speech, and the duration of the phoneme as the feature amount of the speech.
The emotion estimator generation method according to any one of appendices 2 to 5, characterized in that:

(Appendix 7)
In the analysis interval setting step, the speech data is divided into morphemes which are the smallest units having language meaning, and an accent phrase interval combined with a particle or an auxiliary verb spoken after the morpheme is set as the analysis interval. To
The emotion estimator generation method according to any one of supplementary notes 1 to 6, characterized in that:

(Appendix 8)
In the mora section extraction step, when the voice data is displayed as text, one kana character is set as one mora section, the small kana character is combined with the preceding kana character as one mora section, and the long sound is independent. 1 mora section,
The method for generating an emotion estimator according to Supplementary Note 2, wherein:

(Appendix 9)
The emotion estimator estimates the emotion at the time of speaking of the speaker as one of sadness, boredom, anger, surprise, discouragement, disgust, joy,
The emotion estimator generation method according to any one of appendices 1 to 8, characterized in that:

(Appendix 10)
Including a change pattern setting step for setting change patterns classified into the plurality of classes.
The emotion estimator generation method according to any one of supplementary notes 1 to 9, wherein:

(Appendix 11)
An analysis interval setting means for setting an analysis interval for analyzing a feature amount of voice data that is a source of teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Means,
An emotion estimator generator including:

(Appendix 12)
Analysis interval setting means for setting an analysis interval for analyzing features of speech data that is a source of teacher data on a computer;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data means,
Program to function as.

(Appendix 13)
An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation step;
Emotion estimation method including

(Appendix 14)
Analysis interval setting means for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation means;
Emotion estimation device with

DESCRIPTION OF SYMBOLS 1 ... Control part, 2 ... Memory | storage part, 3 ... Input / output part, 4 ... Bus, 100 ... Emotion estimator production | generation apparatus, 110, 210 ... Audio | voice data acquisition part, 120, 230 ... Analysis area setting part, 121, 231 ... Morphological analysis unit, 122, 232 ... Accent phrase extraction unit, 130, 240 ... Accent type determination unit, 131, 241 ... Mora section extraction unit, 132, 242 ... Accent type extraction unit, 140, 260 ... Feature quantity extraction unit, 150 ... Emotion estimator generation unit, 200 ... Emotion estimation device, 220 ... Speaker dividing unit, 250 ... Selection unit, 270 ... Emotion estimation unit, 280 ... Integration unit

Claims (14)

An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the voice data that is the source of the teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Steps,
An emotion estimator generation method including: The change pattern determining step includes:
A mora section extraction step for dividing the speech data included in the analysis section into mora sections that are syllable units;
The speech data of the analysis section is uttered based on the comparison result of comparing the average value of the feature amount of the speech data in the analysis section and the average value of the feature amount of the speech data in the mora section for each mora section. A change pattern extraction step for extracting a change pattern of the feature amount of the sound that changes for each mora section;
The emotion estimator generation method according to claim 1, wherein: In the change pattern extraction step, the fundamental frequency of speech extracted from speech data is used as the feature amount of speech, and the average fundamental frequency of speech in the analysis interval and the average fundamental frequency of speech in the mora interval are determined for each mora interval. Compared to the above, when the average fundamental frequency of the voice in the mora section is higher than the average fundamental frequency of the voice in the analysis section, High is given, and when it is low, Low is given, and the voice changes to High and Low for each mora section. Extract the change pattern of the feature amount of
The emotion estimator generation method according to claim 2. In the change pattern extraction step, the voice intensity extracted from the voice data is used as the voice feature value, and the average voice intensity in the analysis section and the average voice intensity in the mora section are compared for each mora section. When the average intensity of the voice in the mora section is higher than the average intensity of the voice in the analysis section, High is given, and when the average intensity is low, Low is given, and the change in the feature amount of the voice that changes between High and Low for each mora section Extract patterns,
The emotion estimator generation method according to claim 2. In the change pattern extraction step, the phoneme duration extracted from the speech data is used as the feature amount of speech, the average duration of phonemes in the analysis interval, and the average duration of phonemes in the mora interval, Compared for each mora section, High is given when the average duration of phonemes in the mora section is longer than the average duration of phonemes in the analysis section, Low is given when the phoneme is short, and High for each mora section. Extracting the change pattern of the feature amount of the voice that changes to Low,
The emotion estimator generation method according to claim 2. In the change pattern extraction step, a change pattern of the speech feature value is extracted using at least one of the fundamental frequency of the speech, the strength of the speech, and the duration of the phoneme as the feature amount of the speech.
The emotion estimator generation method according to any one of claims 2 to 5, wherein: In the analysis interval setting step, the speech data is divided into morphemes which are the smallest units having language meaning, and an accent phrase interval combined with a particle or an auxiliary verb spoken after the morpheme is set as the analysis interval. To
The emotion estimator generation method according to any one of claims 1 to 6. In the mora section extraction step, when the voice data is displayed as text, one kana character is set as one mora section, the small kana character is combined with the preceding kana character as one mora section, and the long sound is independent. 1 mora section,
The emotion estimator generation method according to claim 2. The emotion estimator estimates the emotion at the time of speaking of the speaker as one of sadness, boredom, anger, surprise, discouragement, disgust, joy,
The emotion estimator generation method according to any one of claims 1 to 8, wherein: Including a change pattern setting step for setting change patterns classified into the plurality of classes.
The emotion estimator generation method according to any one of claims 1 to 9, wherein: An analysis interval setting means for setting an analysis interval for analyzing a feature amount of voice data that is a source of teacher data;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data Means,
An emotion estimator generator including: Analysis interval setting means for setting an analysis interval for analyzing features of speech data that is a source of teacher data on a computer;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. means,
Generation of an emotion estimator for generating an emotion estimator for estimating an emotion of a speaker when speech is uttered for each feature value change pattern, using the voice data classified for each feature value change pattern as teacher data means,
Program to function as. An analysis interval setting step for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Steps,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation step;
Emotion estimation method including Analysis interval setting means for setting an analysis interval for analyzing the feature amount of the audio data to be analyzed;
Change pattern determination for determining a pattern in which the feature amount of the voice data included in the analysis section changes as a pattern of change in the feature amount of the voice data included in the analysis section, based on the change patterns classified into a plurality of classes. Means,
Emotion that estimates the emotion of the speaker when the speech of the analysis section is uttered using an emotion estimator generated based on teacher data having the same feature amount change pattern for each feature amount change pattern An estimation means;
Emotion estimation device with

Images