JP2000181896A - Learning type interaction device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a learning type interaction device capable of learning the association between perception and movement while autonomously performing interaction. SOLUTION: A classifying part 8 classifies a speech pattern inputted from a microphone 1 being a perceptual sensor while dynamically generating a class. An associating part 9 performs association between classes. A sufficiency degree calculating part 10 calculates sufficiency degrees. An expectation generating part 11 searches for an expectation class based on them. An activity calculating part 12 decides whether or not to perform voluntary movement, and a speech searching part 13 searches for a class that should make a speech. Voice being an optimum movement pattern is outputted from a speaker 2 being a movement actuator through a speech generating part 14 and a waveform generating part 15.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning-type interaction device, and more particularly to a learning-type interaction device which autonomously recognizes an external situation and takes an optimal action to perform a spontaneous interaction. It has a function that can learn.

[0002]

2. Description of the Related Art Hitherto, there has been an interactive device for decoding the meaning of a speaker's utterance and generating an optimum response from the meaning.

[0003]

However, in the conventional interactive device, the class of the pattern to be perceived is given in advance at the time of design, and the optimum behavior pattern for the class is also given in advance. That is, there is a problem in that the class classification for the pattern is predetermined by the designer (fixed), and other types cannot respond.

[0004] In addition, the spontaneous operation of the system is merely a system-driven spontaneous utterance of inquiring about an empty slot that lacks knowledge in a conventional interactive device.

[0005] Furthermore, in a system that interacts with humans, the generation of emotional actions required to impart humanity or creatures and to form empathic interactions is performed by anger and sadness, which are called emotional models. It was described as a state transition, and the mapping between emotion and action was also described in advance.

Accordingly, the present invention has been made to solve such a problem, and it is possible to autonomously obtain an optimal action selection through interaction and generate a spontaneous action or an emotional action. It is an object to provide a simple learning type interaction device.

[0007]

According to a first aspect of the present invention, there is provided a learning type interaction apparatus, comprising: a sensing unit for perceiving an external world; a classifying unit for classifying a perceived pattern while dynamically generating a class; Means for associating the classified classes and changing the weight of association to select an optimal output pattern for the perceived pattern, and an actuator for outputting the optimal output pattern for the perceived pattern to the outside world And

[0008] The learning-type interaction device according to claim 2 is the learning-type interaction device according to claim 1, further comprising an expectation means, wherein the expectation means includes means for expecting a perceived pattern, and expectation. Means for calculating a shift between the set pattern and the pattern actually perceived by the sensing means, and means for selecting a strategy for generating an expectation based on an internal state based on the shift.

A learning interaction device according to claim 3 is the learning interaction device according to claim 1 or 2, wherein the sensing means analyzes a microphone that perceives the outside world and a voice, The actuator includes a voice analysis unit that outputs a parameter obtained as a result of the analysis to the classification unit, the actuator includes a speaker that outputs to the outside world, and a voice generation unit that generates a voice that is an optimal output pattern, and a dialogue by voice is performed. Generates a specific language system.

A learning-type interaction device according to a fourth aspect is the learning-type interaction device according to the third aspect, wherein the voice analysis means analyzes the power of the voice, and the voice generation means generates the voice of the analyzed voice. By smoothly changing the power and the frequency, a highly natural sound is generated.

A learning interaction device according to a fifth aspect is the learning interaction device according to the third aspect, wherein the voice analyzing means converts the voice perceived from the outside world and the voice generated by the voice generating means. analyse.

[0012]

[First Embodiment] A learning type interaction device according to a first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing an outline of an overall configuration of a learning interaction device according to Embodiment 1 of the present invention. Referring to FIG. 1, a learning type interaction device includes a microphone 1, a speaker 2, an A / D converter 3, a power calculator 4,
It includes a voice section detection unit 5, a normalization unit 6, a distance calculation unit 7, a classification unit 8, and an association unit 9.

The learning interaction apparatus according to the first embodiment of the present invention is a voice interactive system including a microphone 1 as a sensor for perception and a speaker 2 as an actuator for action. The audio signal input from the microphone 1 is
In the A / D converter 3, the data is digitized at 16 kHz and 16 bits.

In the power calculation unit 4, 512 samples (1
The dialog voice power is calculated for each frame shift.
Assuming that the frame length is L, the dialog voice power logp is
It is represented by equation (1).

[0015]

(Equation 1)

For example, if the frame length is L = 512,
Alternatively, the calculation is performed using L = 768. In order to remove the fluctuation of the background sound, the level is raised by equation (2).

[0017]

(Equation 2)

For example, the backgroundroundve
1 = 30 (dB). The voice section detection unit 5 calculates an average power for each frame, and determines that voice has been input if the difference from the average power exceeds a certain value. Average power a
Vepow is updated by Expression (3).

[0019]

(Equation 3)

However, this indicates that the sound is off. When the sound is on, the on-time (spcnt) of the sound and the time constant MAX_UTTER_CNT_ARTICULA of the sound are used due to the effect of the long sound.
Based on the TE and the average power avep using equation (4)
Find ow.

[0021]

(Equation 4)

When a state in which the value of (logp-avepow) is larger than a predetermined number of THONs continues for a predetermined number of frames CNTON, the state is turned on, and the predetermined number THO is set.
If the state smaller than the FF continues for the predetermined number of frames CNTOFF, the state is set to the off state.

The voice section detecting section 5 receives data of power spoken by the system at the same time as voice input from the outside. As a result, the utterance data uttered by itself and the utterance data input from the outside are equivalently segmented (speech segmentation), and the subsequent processing can be treated equivalently.

The voice section detection section 5 outputs an on / off state of voice input and a time-series pattern of voice power when the voice is turned off. Note that external voice data is classified by attaching a flag ch = 0, and internal utterances by a flag ch = 1.

When the distance between the microphone 1 and the speaker 2 is short, the speech of the system enters the microphone 1. To avoid this, the power of the utterance of the system is subtracted when calculating the power. Log conversation power of a certain frame t
p (t), the utterance power of the system as OutPower
Assuming (t), the dialog power logp (t) is expressed by equation (5).

[0026]

(Equation 5)

Here, the coefficient a (i) and the number of characters N represent equivalent transfer characteristics from output to input. For example, a (1) = 1.0, a (2) = 1.0, and N = 2.

The normalization unit 6 cuts out the last part of the utterance by weighting the longest length accepted by the system for an utterance that is too long, and pads the last part of the utterance with zeros for an utterance that is too short. And adjust the utterance length to the minimum length accepted by the system.

The distance calculator 7 calculates the distance between the input voice pattern and the pattern class stored in the system, and obtains the minimum distance class. Here, a pattern class included in the system will be described with reference to FIG. FIG. 2 is a diagram showing a hierarchical structure when classifying the pattern classes. Referring to FIG. 2, the learning interaction device includes a class having a plurality of hierarchies, and each class has a tree structure. In FIG. 2, for example, 3
This shows the hierarchy.

The first hierarchy N1 includes a class C1 for classifying by time (hereinafter, referred to as a time class). The second layer N2 is a class C for performing a large classification of the pattern.
2 (hereinafter referred to as major class). Third level N
No. 3 includes a class C3 for performing finer classification of patterns (hereinafter, referred to as a minor class). Each of the second hierarchy (major class) and the third hierarchy (minor class) has a class representative pattern representing a class.

The patterns are first classified by the time class C1. Specifically, classification is performed based on the pattern length. For example, 16 time classes C1 are prepared, and each time class C1 is prepared.
Range is 4 frames. 15 from 12 frames in length
The time class C1 (0) up to the frame, the time class C1 (1) from the 16th frame to the 19th frame,
.., And the frames from 72 to 75 are defined as a clock class C1 (15). Anything that deviates from this range
It is normalized by the normalization unit 6.

After being classified into a certain time class C1, the second
The distance to the major class C2 of the hierarchy N2 is calculated.
The distance D is represented by Expression (6), where x (t) is the input pattern and c (t) is the class representative pattern.

[0033]

(Equation 6)

Here, L represents the pattern length. The number of terminals of the input pattern length is padded with zeros. Using Expression (6), the input pattern is classified into the major class C2 having the smallest distance in the second hierarchy. Then, the distance calculation is similarly performed on the third class minor class C3, and the minimum distance minor class C3 is obtained. The classification error at this time is denoted as Dmin.

Referring to FIG. 1, the classification unit 8 generates a new class when the classification error Dmin is larger than the class generation criterion of the major class and the minor class, and otherwise generates the minimum distance. Classification into classes.

Specifically, when the classification error Dmin is larger than the distance MjD, which is a measure for generating a major class, a new major class representing the input pattern is generated, and the input pattern is placed under the new major class. And generate a minor class. Then, the input pattern is classified into this class.

On the other hand, if the classification error Dmin is equal to the distance MjD
If it is smaller, the distance is compared with the distance MrD, which is the generation reference of the minor class. The classification error Dmin is equal to the distance M
If it is greater than rD, create a new minor class. Then, the input pattern is classified into this class.

When a new minor class is generated, a related link is generated for the own class. Further, an accumulated classification error is stored in each minor class, and the classification errors are accumulated when the input pattern is classified into that class. Cumulative classification error is a certain value MrC
If it is larger than umD, a new minor class is generated. Furthermore, a new minor class is also generated when the number of related lists possessed by the minor class is larger than a certain value MRCLASS_COMPLEXITY. Each threshold value (reference value) shown here is not a fixed value but can be changed in a learning situation.

The associating unit 9 associates the input utterance classes. The minor class has a weighted link for itself and a weighted link for other minor classes. This association is the association of the utterance act with the perception.

The association of the utterance act with the perception will be described with reference to FIG. FIG. 3 is a diagram for explaining the association of the utterance act with the perception. Referring to FIG. 3, an utterance and a causal relationship between the utterances are regarded as a class association, and a link is generated. FIG. 3 shows the association between the classes 2 and 3 relating to the utterance of the user and the class 3 possessed by the system. The symbol interval is
Indicates the interval between utterances. For example, as shown in FIG. 3, when the utterance of the user is classified into class 3, a link is generated from class 2 to class 3, or the link weight is adjusted.

Specifically, a link for class 3 in class 2 is searched for. If there is no link, create a new link. New link weight RelEnergy
(Hereinafter referred to as RE) is the interval interva between utterances.
l, maximum time to associate TRel, and update factor R
From elAlfa, it is determined based on equations (7) and (8).

[0042]

(Equation 7)

[0043]

(Equation 8)

A link to another class uses equation (9).

[0045]

(Equation 9)

As described above, by reducing the weight RE using the equation (9), the sum of the weights for all the links is normalized to 1.0.

When a new class is generated, a weighted link of 1.0 is generated for its own class. If links already exist, the weight RE is set to equation (10) for all links.

[0048]

(Equation 10)

Further, for a link to class 3, equation (11) is used.

[0050]

[Equation 11]

Thus, the weight is increased while normalizing. When a new class is created, a link to itself (self-link) is created. At this time, since the self-link destination class is selected as the optimal action, a simulated response to an external input can be generated.

Further, when the association from class 2 to class 3 is generated, a reverse link in the reverse direction from class 3 to class 2 is also generated at the same time. In this case, the value TRatio is set to the value In.
multiply by factorFactor. For example, the value Inv
Inverse factor = 0.1, and if no reverse cause is given, Inverse factor = 0.0.

Referring to FIG. 1, the learning interaction apparatus further includes a sufficiency calculation unit 10, an expectation generation unit 11, an activity calculation unit 12, an utterance search unit 13, an utterance generation unit 14, and a waveform generation unit 15. Prepare.

The sufficiency calculation unit 10 calculates the sufficiency F as an internal parameter and the average sufficiency Fav which is a temporal average thereof.
and e. These play an important role in choosing a strategy of action choice. The sufficiency F is calculated by the difference between the generated expectation and the actual input. Satisfaction degree F
Is defined by equation (12).

[0055]

(Equation 12)

Here, the value M is the degree of consistency between the input class and the expectation, the value L is the difference between the input class and the expectation, and the value Pg
Is the shift for distances similar to the pattern of the input class,
The value Tg represents a time lag until an input with respect to an expectation comes.

Actually, the sum is obtained by weighting according to the range of the value of each element. Here, the sufficiency F is a one-dimensional value as an example, but may be multi-dimensional as an internal space.

When there is an input event, that is, when there is an utterance from the outside or the system itself, the classification error Dmin to the input class is substituted as the value Pg. Next, it is searched whether or not the input class exists in the expectation generated by the expectation generation unit 11 described later. If the class input in the expectation exists, and the probability of the expectation of the class is CE, the value M is updated using the equation (13), and the value L is updated using the equation (14). I do.

[0059]

(Equation 13)

[0060]

[Equation 14]

On the other hand, if the input class does not exist in the expectation, the value M is adjusted by the equation (15) and the value L is adjusted by the equation (16).

[0062]

(Equation 15)

[0063]

(Equation 16)

To prevent the values M and L from diverging, the value Mde
c, Limit the maximum value using Link. Then, the sufficiency F is calculated from these values using Expression (12).

Even if there is no event, the sufficiency F is updated for each frame. The value Tg is increased based on Expression (17) unless an utterance is input.

[0066]

[Equation 17]

When an utterance is input, the value Tg = 0. Except for the value Tg, as shown in Expression (18), Expression (19), and Expression (20), the shift is attenuated over time by decreasing the weight of each factor.

[0068]

(Equation 18)

[0069]

[Equation 19]

[0070]

(Equation 20)

Here, the average degree of satisfaction Fave, which is the temporal average of the degree of satisfaction F, is calculated by, for example, equation (21).

[0072]

(Equation 21)

The emotion degree Emotive used in the activity calculator 12 described later is calculated based on the equation (22).

[0074]

(Equation 22)

The degree of emotion Emotive is calculated as the sum of the factors that evoke emotion and used as a criterion for taking action. The average satisfaction level Fave is a smooth parameter for temporal simplification corresponding to mood in emotion. In the first embodiment, since the value Tg is treated as an independent action source, the average degree of satisfaction Fave is not included in the calculation of the emotional degree Emotive.

When there is an utterance event, the expectation generator 11 generates an expectation according to the values of the degree of satisfaction F and the average degree of satisfaction Fave. These strategies become strategies of anticipation and action, creating emotional behavior.

There are two types of expectation, an expectation for an utterance input from the outside and an expectation for an utterance of the user.
The expectation of the utterance input from the outside is an expectation of an action to be performed next by the system. The expectation of the user's own utterance is the expectation of the next class to be uttered by an external user.

In the generation of the expectation, the probabilities are calculated from the class serving as the starting point (the input class) by the related link, and the number of expectation classes according to the situation is searched. Here, the processing of the expectation generation unit 1 will be described with reference to FIG. FIG. 4 is a diagram for explaining calculation for generating an expectation, and shows a state of a link between the classes in the classes 1 to 10. In FIG. 4, there are cases where there is a self-link other than the class 1 but it is abbreviated.

Referring to FIG. 4, each class 1 to 10 has expectation probabilities CE1 to CE10, and is connected to another class or itself through link weights RE1 to RE10 generated by association.

For example, for class 1, four primary links of class 1, class 2, class 3 and class 4 are connected by links. Thus, when calculating expectations starting from class 1 having an expectation probability CE1, a new expectation probability for each is calculated according to: For example, the expected probability CE1 is calculated based on Expression (23).

[0081]

(Equation 23)

The expected probability CE2 for class 2
Is calculated based on equation (24).

[0083]

(Equation 24)

Here, the coefficient alfa represents the weight for the current expected probability CE, and the coefficient beta represents the weight for the new link. Note that (alfa + bet
a) = 1, and the expectation is likely to shift when the coefficient beta is large.

A new expected probability CE is calculated, the values are sorted from the largest one, and the classes are selected by the number determined from the top one. Then, normalization is performed so that the sum of the expected probabilities CE becomes 1. If the number of links is small, calculation is performed up to the secondary link. In that case, the search is performed from the primary link having the highest expected probability CE.
In the case of FIG. 4, the range enclosed by the broken line is the calculation range,
Calculation is performed for class 3 up to the secondary link.
As described above, the expectation is calculated by the number determined by the number of expected probabilities CE or the limit of the degree of the link.

The expected probabilities CE of the other classes are 0.
And Since the calculation of the expectation is performed only for the required number of classes, the amount of calculation is small even if the number of classes increases. Also, by adaptively determining a trade-off between the amount of calculation proportional to the specified number and the response of the system, it is possible to characterize the responsiveness and contemplation of the system.

Satisfaction degree F and average satisfaction degree F as internal states
By selecting a strategy for generating an expectation with ave, the expectation and the action selected with it can be accurately identified. As an example, the satisfaction degree F and the average satisfaction degree Fav
In combination with e, an action strategy is selected as shown in the following specific example.

When the satisfaction degree F ≧ 0 and the average satisfaction degree Fave ≧ 0, both the current satisfaction degree and the time-average satisfaction degree are positive values, indicating that the prediction is very successful. In this case, even if the range of the expectation is narrowed, it is OK, and the system can respond quickly.

When the satisfaction degree F ≧ 0 and the average satisfaction degree Fave <0, the expectation has not been successful for a while, but the present expectation has been successful, so that the expectation is broadly generated.

When the degree of satisfaction F <0 and the average degree of satisfaction Fave ≧ 0, the expectation has been successful for a while, but the present expectation has been missed. By maintaining the conventional expectations without altering the expectations, this is equivalent to respecting the internal expectations, which is an act of feeling self-asserted anger.

When the satisfaction degree F <0 and the average satisfaction degree Fave <0, in the expectation of the utterance input from the outside,
An expectation is generated so as to increase the weight of a class input from the outside, and this class is selected by the utterance search unit 13, thereby imitating an opponent. Thus, it appears to be obedient. Represents a state close to sadness. In addition, in the expectation of the user's utterance, the user removes his / her own class so as not to make the same utterance and generates the expectation, thereby generating a variety of actions. Where the average degree of satisfaction Fav
e can also be grasped as a parameter representing comfort / discomfort.

Referring to FIG. 1, activity calculating section 12 determines whether or not to take action based on an external event or expectation. This is set as the true / false flag.

First, when an external event is input,
The flag Action is true. This is equivalent to always responding if there is any input as a system. If the absolute value of the average satisfaction level Fave exceeds a certain value, the flag Action is set to true. The average satisfaction level Fave is a temporally smooth parameter corresponding to the mood in emotion as described above, and indicates that when it is working well, when the gap is large, or when it performs spontaneous action according to the situation. Means. Further, when the emotion degree Emotive exceeds a certain value, the flag Actio
Let n be true. This corresponds to reacting when an event occurs in the short term.

If the value Tg exceeds a certain value, the flag A
Let ction be true. This means that the system prompts for an action when there is no response from the other party for a certain period of time. Average Satisfaction Fave and Emotion Emotiv
When the flag Action becomes true once by e, the threshold value for determining the flag Action is increased for a while to avoid that the flag Action becomes true twice. The increased threshold is reduced in time.

The utterance searching unit 13 searches for a class to be uttered when the activity calculation unit 12 sets the flag Action to true. Specifically, the class having the highest expected probability CE is searched from the expected classes.

The utterance generation unit 14 generates a time series of audio power for each frame from the pattern of the class to be uttered. This output is input to the voice section detection unit 5.

The waveform generator 15 generates a waveform from the audio power data. As an example, an algorithm for generating a waveform from power data p (1),..., P (N) of N consecutive frames is described below. In a certain frame, waveform sample data of only frame shift (FRSHIFT) is generated. To increase the naturalness of the waveform, the frequency is increased in proportion to the increase in power. Also, the power and frequency are smoothly changed for naturalness. The waveform is, for example, 1 time of a sine wave, 2 times.
The sum of the double and triple components is generated.

First, power data p of a certain frame
(I) From (1 ≦ i ≦ N), the value De is calculated based on equation (25).
Calculate itaPower.

[0099]

(Equation 25)

The frequency Freq (i) of a certain frame is
It is calculated based on equation (26).

[0101]

(Equation 26)

Here, BaseFreq is a constant indicating the reference frequency, and freq_sense is a constant indicating the degree to which the frequency fluctuation is determined by the power data p (i).

In order to smooth the frequency change, the frame frequency Freq (i) is smoothed using the equation (27).

[0104]

[Equation 27]

Next, using the result of equation (27), equation (2)
The value DeltaFreq is determined based on 8).

[0106]

[Equation 28]

Waveform sample x (t) (where t = 0,
..., FRSHIFT-1) sets the sampling frequency to F
Let S be the phase phase (t) shown in equation (29).
, And the amplitude amp (t) shown in the equation (30).

[0108]

(Equation 29)

[0109]

[Equation 30]

[0110]

(Equation 31)

The waveform can be generated with reference to a waveform table. Based on the above description, the process of the learning interaction apparatus according to Embodiment 1 of the present invention will be described with reference to FIGS. 5 and 6
5 is a flowchart for explaining a process performed by the learning interaction apparatus according to Embodiment 1 of the present invention.

Referring to FIG. 5 and FIG. 6, step S1
Then, the power calculator 4 calculates the power of the input voice (one frame). In step S2, the voice section detection section 5 detects a voice section of the input voice and the self-utterance. In step S3, it is determined whether an utterance has been detected. If no utterance has been detected, the process proceeds to step S9 described below.

In the subsequent step S4, the normalizing section 6 normalizes the utterance. In step S5, the distance calculator 7 calculates the distance between the class closest to the utterance pattern and the utterance pattern.

In the following step S6, the categorizing section 8 compares the obtained distance with the class generation standard. If the distance is larger than the class generation standard, the process proceeds to step S7, where a new class is generated. If it is less than the class creation criteria, and after a new class is created,
In step S8, the associating unit 9 associates the utterance classes.

In the following step S9, the sufficiency calculation section 10
In, the calculation (update) of the sufficiency is performed. In the following step S10, if there is a speech event, the process proceeds to step S11. In step S11, the expectation generation unit 11 generates an expectation. When there is no utterance event and after the generation of the expectation, the process proceeds to step S12, and the activity calculation unit 12 calculates the activity (flag Action). Step S
At 13, it is determined whether the flag Action is true. If the flag Action is true, the process proceeds to step S14, and the utterance search unit 13 searches for a class to be uttered. Start speaking yourself.

If the flag Action is not true and after the utterance search, it is determined in step S15 whether or not the own utterance has ended. If the processing has not been completed, the process proceeds to step S16, where a waveform is generated via the utterance generation unit 14 and the waveform generation unit 15 and output from the speaker 2 (for one frame). After the end of step S15 or step S16, the process returns to step S1.

Each parameter shown in the first embodiment of the present invention is one example, and the present invention is not limited to this. It can be changed appropriately in specific embodiments. In addition, the sensor to be handled is not limited to a voice microphone, but can be perceived as multidimensional sensory data using a camera or a tactile sensor, and an action can be generated by a display or a robot arm.

The embodiments disclosed this time are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

[0119]

As described above, according to the present invention, it is possible to acquire patterns of perception and action while interacting autonomously without giving a designer a method of generating perception and actions for perception in advance. In addition, an interaction system that learns the association between perception and action can be configured.

In addition, a mechanism for generating a spontaneous action performs a spontaneous interaction,
More learning can be done. In addition, since the strategy of spontaneous action generation generates actions similar to those based on human emotions, the system becomes easy for humans to understand when interacting with humans.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an overall configuration of a learning interaction device according to Embodiment 1 of the present invention.

FIG. 2 is a diagram showing a hierarchical structure when pattern classes are classified.

FIG. 3 is a diagram for explaining association of an utterance act with respect to perception;

FIG. 4 is a conceptual diagram illustrating a calculation for generating an expectation.

FIG. 5 is a flowchart illustrating a process of the learning interaction apparatus according to the first embodiment of the present invention.

FIG. 6 is a flowchart for explaining a processing process of the learning interaction apparatus according to the first embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 microphone 2 speaker 3 A / D conversion unit 4 power calculation unit 5 voice section detection unit 6 normalization unit 7 distance calculation unit 8 classification unit 9 association unit 10 sufficiency calculation unit 11 expectation generation unit 12 activity calculation unit 13 utterance search Unit 14 utterance generation unit 15 waveform generation unit

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] September 28, 1999 (1999.9.2)
8)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claim 2

[Correction method] Change

[Correction contents]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0008

[Correction method] Change

[Correction contents]

[0008] A learning-type interaction device according to a second aspect is the learning-type interaction device according to the first aspect, further comprising an expectation unit, and the expectation unit is notified next by the sensing unit.
Calculate the probabilities of the class in which the perceived pattern is classified
Means that, based on the calculated probabilities, then Pa perceived
Means for calculating a deviation between the actual perceived pattern belongs class by class and sensing means turns are categorization, the internal state based on the shift to the original, then perceived
Means for calculating the probability of the class in which the pattern to be classified is categorized .

Continued on the front page F term (reference) 5H004 GA26 HA20 HB15 JB18 JB19 KD63 LA17 9A001 BB04 DD13 EE05 FF10 FZ03 GZ05 HH05 HH16 HH18 HH21 JZ06 KZ32

Claims (5)

[Claims] 1. A sensing means for perceiving the outside world, a classifying means for classifying the perceived pattern while dynamically generating a class, and an association between the classified classes is provided. A learning type interaction device comprising: means for changing a weight to select an optimal output pattern for the perceived pattern; and an actuator for outputting the optimal output pattern for the perceived pattern to the outside world. 2. The apparatus according to claim 1, further comprising: an expectation unit, wherein the expectation unit expects the perceived pattern, and calculates a deviation between the expected pattern and the pattern actually perceived by the sensing unit. The learning-type interaction device according to claim 1, further comprising: a means for selecting a strategy for generating the expectation based on an internal state based on the deviation. 3. The sensing unit includes: a microphone that perceives the outside world; and a voice analysis unit that analyzes a voice and outputs a parameter obtained as a result of the analysis to the classification unit. The speaker according to claim 1 or 2, further comprising: a speaker that outputs to the outside world; and a voice generation unit that generates voice that is the optimal output pattern, wherein a system of a specific language is generated by the dialogue with the voice. Learning type interaction device. 4. The voice analyzing means analyzes the power of the voice, and the voice generating means generates a highly natural voice by smoothly changing the power and frequency of the analyzed voice. The learning-type interaction device according to claim 3. 5. The learning interaction apparatus according to claim 3, wherein the voice analysis unit analyzes the voice perceived from the outside world and the voice generated by the voice generation unit.