JP6111802B2 - Spoken dialogue apparatus and dialogue control method - Google Patents

Description

  The present invention relates to a voice dialogue apparatus and a dialogue control method for recognizing a voice uttered by a user and executing processing according to the recognition result, for example.

  In recent years, a voice interaction system has been developed in which a voice signal input via a terminal is transmitted to a server via a communication network, the server selects desired content according to the voice signal, and the content is returned to the terminal. ing. In such a spoken dialogue system, when the load on the server exceeds a certain level, or when the communication network is congested, the delay in returning information from the server to the terminal may increase. there were. When the delay increases, the response time during which the terminal does not respond increases from the voice input by the user until the terminal presents the content corresponding to the voice to the user. For this reason, the user feels anxiety or the user's discomfort increases, and as a result, the user's convenience for the voice interaction system may be reduced. In view of this, there has been proposed a voice interaction device that predicts the time required for reading the recognition dictionary based on the size information of the recognition dictionary and outputs a response voice having a length corresponding to the predicted time to a speaker (for example, , See Patent Document 1). This spoken dialogue apparatus prevents such a long silence period from being read when the recognition dictionary is read out by outputting such a response voice.

JP 2001-22384 A

  However, the response time varies depending on the communication state of the communication line between the terminal and the server or the processing load of the server. For this reason, the response time is not accurately estimated only by the size information of the recognition dictionary, and the message length may not be optimized.

  Therefore, according to one aspect of the present invention, even if the response time varies, the length of the message voice to be played during the response time can be set so that the deterioration of the user's quality of experience during the response time can be suppressed. An object is to provide an interactive device.

  According to one embodiment, a voice interaction device is provided. This voice interactive apparatus includes a voice recognition unit that extracts a predetermined keyword from a voice signal representing a user's voice collected by the voice input unit, a search unit that searches for content according to the keyword, and a user's voice input A delay distribution calculation unit that calculates a probability distribution of response time from the end of the time until the content is presented to the user, and an estimated response time or an estimation when the response time is less than or equal to the estimated response time based on the probability distribution The first expected value for the first experience quality of the user when the first message voice having a length obtained by subtracting a predetermined offset from the response time is reproduced within the response time, and the response time than the estimated response time Is the sum of the second expected value for the second experience quality of the user when the first message voice and the additional second message voice are reproduced within the response time. A message length determination unit for determining the length of the message voice so as to maximize the value, a message generation unit for generating text information of the first message having the determined message voice length, and a first And a voice synthesizer that synthesizes the first message voice from the text information of the message.

The objects and advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims.
It should be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are not restrictive of the invention as claimed.

  The voice interactive device disclosed in this specification sets the length of the message voice to be played during the response time so that the deterioration of the user's quality of experience during the response time can be suppressed even if the response time varies. it can.

1 is a schematic configuration diagram of a voice interaction system according to one embodiment. It is a functional block diagram of the control part of a server. It is a figure which shows an example of an information source table. It is a figure showing the change by the response time of a user's bodily sensation quality when the estimation of a response time is successful, and a user's bodily sensation quality when estimation of a response time fails. It is an operation | movement sequence diagram of a dialog control process. (A) is a figure showing an example of the function showing the relationship between response time and bodily sensation quality. (B) is a figure showing an example of the cumulative probability of response time. (C) is a figure showing the relationship between the cumulative probability of response time and the user's quality of sensation obtained from (a) and (b). In the graph showing the relationship between the cumulative probability P and the sensory quality shown in FIG. 6C, the sensory quality R (P) is a constant value R, and the relationship between the cumulative probability P and the sensory quality Q (P) is It is a figure showing the graph approximated by the straight line. It is a figure showing the relationship of each parameter used for calculation of the cumulative probability that unextended experience quality and extended experience quality correspond.

Hereinafter, a voice interaction device and a dialogue control method of the voice interaction device according to various embodiments will be described with reference to the drawings.
This voice interaction apparatus estimates a probability distribution of response time from when a user inputs voice through the terminal until the terminal responds to the inputted voice based on past response time. The voice interaction apparatus determines the length of the message to be reproduced during the response time based on the probability distribution of the response time so that the user's quality of experience is the highest. At that time, this voice interactive device sends an additional message when the actual response time becomes longer than the estimated value so that the quality of experience can be statistically improved even if the response time varies. Consider the quality of experience when inserting.

  FIG. 1 is a schematic configuration diagram of a voice interaction system. In this embodiment, the voice interaction system 1 selects a terminal 2 for receiving content provided according to the input voice by the user, and a content corresponding to the voice input via the terminal 2. And a server 3 for returning the selected content to the terminal 2.

  The terminal 2 and the server 3 can communicate with each other via a communication network 4 such as a public communication line. Further, the server 3 may be able to communicate with an external information source 5 such as a Web server or an ftp server via the communication network 4. Note that the terminal 2 and the server 3 may be connected to a communication line such as a dedicated line separate from the communication network 4.

  The terminal 2 is, for example, a mobile phone, a portable information terminal, or a fixed terminal, and includes a voice input unit 21, a voice output unit 22, a communication unit 23, a storage unit 24, and a processing unit 25. Further, the terminal 2 may further include a display unit (not shown) such as a liquid crystal display.

  The voice input unit 21 includes, for example, a microphone and an analog / digital converter. The voice uttered by the user is converted into an audio signal which is an analog electric signal by the microphone. The analog audio signal is sampled at a predetermined sampling frequency by an analog / digital converter, converted into a digital audio signal, and then sent to the processing unit 25.

  The audio output unit 22 is an example of an output unit that presents content to the user, and includes, for example, a digital / analog converter and a speaker. Then, the audio output unit 22 converts the audio signal received from the processing unit 25 into an analog signal by a digital / analog converter, and the analog audio signal is converted into a sound by a speaker and output to a user.

  The communication unit 23 includes an interface circuit for connecting the terminal 2 to the communication network 4. And the communication part 23 transmits the audio | voice signal received from the process part 25 to the server 3 via the communication network 4 according to the communication system with which the communication network 4 is based. On the other hand, the communication unit 23 receives the content and the message voice reproduced during the response time from the server 3 via the communication network 4 and passes them to the processing unit 25.

  The storage unit 24 includes, for example, a semiconductor memory. And the memory | storage part 24 memorize | stores the various programs for the process part 25 to control the terminal 2, the application program which operate | moves on the terminal 2, and various data required for execution of these programs.

  The processing unit 25 includes one or more processors and peripheral circuits. Then, the processing unit 25 outputs the audio signal input via the audio input unit 21 to the communication unit 23. When the voice input by the user is completed, the processing unit 25 reproduces the message voice received from the server 3 via the voice output unit 22. Note that the processing unit 25 obtains a power spectrum for each frame by, for example, performing time-frequency conversion on the input audio signal in units of predetermined frames in order to determine whether or not the audio input has been completed. Then, the processing unit 25 determines that the voice input is completed when a predetermined number or more of frames having the average power value or less continue after the frame in which the average power value for each frequency band is equal to or higher than the predetermined threshold value. To do.

  Further, when the processing unit 25 receives the content from the server 3 after the reproduction of the message sound is completed, the processing unit 25 reproduces the sound signal included in the content via the sound output unit 22.

If the content includes an image or video stream, the processing unit 25 may display the image or video stream on a display unit (not shown). Further, the processing unit 25 may cause the display unit to display the text information even when the content includes text information.
Furthermore, when the application program such as a game is included in the content, the processing unit 25 executes the application program, causes the display unit to display an image corresponding to the execution result, or outputs an audio signal corresponding to the execution result. The output unit 22 may reproduce the data.

  Further, the processing unit 25 measures the time from the end of voice input until the content is received as an actual response time, and transmits the response time to the server 3 via the communication unit 23.

  Referring again to FIG. 1, the server 3, which is an example of a voice interaction apparatus, includes a communication unit 31, a storage unit 32, and a control unit 33.

The communication unit 31 has an interface circuit for connecting the server 3 to the communication network 4. Then, the communication unit 31 passes the audio signal received from the terminal 2 via the communication network 4 to the control unit 33 in accordance with the communication method that the communication network 4 complies with. Further, the communication unit 31 stores the response time received from the terminal 2 in the storage unit 32. On the other hand, the communication unit 31 transmits the content, message voice, and the like received from the control unit 33 to the terminal 2 via the communication network 4. Further, the communication unit 31 may transmit an information acquisition request signal to the external information source 5 via the communication network 4 and receive data including predetermined content from the external information source 5.
Hereinafter, for convenience of explanation, the user's voice signal received by the server from the terminal 2 is referred to as an input voice signal.

  The storage unit 32 includes, for example, at least one of a semiconductor memory, a magnetic recording device, and an optical recording device. And the memory | storage part 32 memorize | stores the various programs for the control part 33 to control the server 3, the application program which operate | moves on the server 3, and various data required for execution of these programs. For example, the storage unit 32 stores various information used for speech recognition or speech synthesis, various contents requested by the user, or the address of the external information source 5. Further, the storage unit 32 stores various types of information used for determining the length of the message voice reproduced during the response time and the content of the message voice. For example, the storage unit 32 stores a probability distribution of response time, a table representing the relationship between the response time and the user's experience quality, and the like.

  The control unit 33 includes one or more processors, a readable / writable semiconductor memory, and a peripheral circuit. And the control part 33 extracts a keyword from an input audio | voice signal, and searches the content which a user desires based on the keyword. Furthermore, the control unit 33 sets the time length of the message voice to be reproduced within the response time according to the probability distribution of the response time, and creates the message voice according to the time length. Then, the control unit 33 returns the searched content and message voice to the terminal 2.

FIG. 2 is a functional block diagram of the control unit 33 of the server 3 included in the voice interaction system. The control unit 33 includes a voice recognition unit 34, a search unit 35, a delay distribution calculation unit 36, a message length determination unit 37, a message generation unit 38, and a voice synthesis unit 39.
Each of these units included in the control unit 33 is, for example, a functional module realized by a computer program that operates on a processor included in the control unit 33. Note that these units included in the control unit 33 may be mounted on the server 3 as one integrated circuit that realizes the functions of the units.

  The voice recognition unit 34 extracts a keyword for content search registered in advance from the input voice signal. In order to extract a keyword, the speech recognition unit 34 may use any of various speech recognition methods such as a method using an acoustic model or a method using a dynamic time stretching method. In the present embodiment, the speech recognition unit 34 extracts a keyword using a word dictionary storing a keyword to be recognized and an acoustic model created in advance.

  The acoustic model is generated, for example, by dividing the pronunciation of a keyword into unit sounds such as phonemes or syllables, and connecting unit acoustic models corresponding to the unit sounds in accordance with the order of the unit sounds. Is remembered. The unit acoustic model and the acoustic model are represented by, for example, a hidden Markov model (HMM). The speech recognition unit 34 uses an HMM representing a unit acoustic model, and based on one or more feature amounts extracted from a predetermined section of the input speech signal, the predetermined section for a specific unit sound is an estimated sound. Probability or likelihood is calculated.

  Specifically, the voice recognition unit 34 extracts feature amounts used for voice recognition from the input voice signal. For this purpose, the speech recognition unit 34 obtains a spectrum for each frame by performing frequency conversion such as fast Fourier transform for each frame having a predetermined frame length on the input speech signal. The frame length is set to, for example, about 10 milliseconds to 100 milliseconds. Based on the spectrum, the speech recognition unit 34 obtains, for example, a Mel Frequency Cepstral Coefficient (MFCC) or a power difference value between frames as a feature amount for each frame. For example, when the MFCC is calculated as the feature amount, the speech recognition unit 34 converts the spectrum of each frame into a mel scale power value, and then performs frequency conversion such as discrete cosine conversion again on the logarithm of the power value. To calculate the MFCC. Further, when obtaining the power difference value between the frames as the feature amount, the voice recognition unit 34 obtains, for example, the sum of the squares of the spectrum for each frequency band of each frame as the power, and the power between the two consecutive frames. The difference value is obtained by obtaining the difference.

  Note that the speech recognition unit 34 may extract any of various other feature amounts (for example, fundamental frequency) used in speech recognition using an acoustic model as the feature amount. The speech recognition unit 34 may extract a plurality of types of feature amounts for each frame from the input speech signal.

  The speech recognition unit 34 collates the acoustic model with a set of feature values obtained from one or more frames, thereby for each keyword included in the word dictionary, the likelihood for each keyword generated by concatenating unit sounds. Find the degree. Then, the voice recognition unit 34 extracts a predetermined number of keywords in descending order of likelihood. The predetermined number is set to about 1 to 5, for example. Then, every time a keyword is detected, the voice recognition unit 34 notifies the search unit 35 and the message generation unit 38 of the keyword.

  The search unit 35 specifies an information source corresponding to the detected keyword by referring to an information source table representing the correspondence relationship between the keyword and the information source. And the search part 35 acquires the content corresponding to a keyword from the information source. The content includes, for example, at least one of text information, image data, and an application program.

  FIG. 3 is a diagram illustrating an example of the information source table. At least one keyword is stored in each column in the left column of the information source table 300. On the other hand, each column in the right column of the information source table 300 stores address information indicating information sources corresponding to the keywords shown in the same row. This address information is, for example, a uniform resource locator (URL) for specifying the external information source 5 or a file name including specific content stored in the storage unit 32. One address information may be associated with a plurality of keywords, and one keyword may be associated with a plurality of address information.

The search unit 35 refers to the information source table and identifies address information corresponding to the input keyword. And the search part 35 accesses the information source shown by address information, and acquires the content corresponding to a keyword. For example, the search unit 35 reads a file having the file name indicated in the address information from the storage unit 32. Alternatively, the search unit 35 receives the web page specified by the URL from the external information source 5 specified by the URL indicated in the address information via the communication network 4. And the search part 35 analyzes the source | sauce of the received web page, and extracts the text information and image information which are displayed on a screen within a web page as a content.
The search unit 35 outputs the content to the communication unit 31. If the text information is included in the content, the search unit 35 passes the text information to the speech synthesizer 39.

  When the control unit 33 receives the input audio signal, the delay distribution calculation unit 36 obtains a frequency distribution for each response time from the response time history of the latest fixed period. Then, the delay distribution calculation unit 36 calculates the occurrence probability for each response time by dividing the frequency for each response time by the total number of response times during the certain period, thereby obtaining the probability distribution of the response time. The certain period is set to, for example, 10 minutes to 1 hour.

In the present embodiment, the server 3 acquires content from an arbitrary Web server or the like via the communication network 4 or transmits the content to the terminal 2 via the communication network 4. Therefore, the response time varies between several hundred milliseconds to several seconds, for example, depending on the communication state of the communication network. The communication state varies depending on, for example, a time zone or time. The response time also varies depending on the processing load of the external information source 5 itself that provides the content. For example, in the night when there are few users, the response time is relatively short and the variation is small. On the other hand, during commuting hours, lunch breaks, or when returning home, which are times when there are many users, the response time is relatively long and the variation is also large. Furthermore, even when a new service is started on a web server related to a popular site, or when the deadline for any application period is about to expire, the response time becomes relatively long and the variation becomes large.
Therefore, as in the present embodiment, the delay distribution calculation unit 36 can improve the accuracy of the probability distribution of the response time by using the history of the response time in the latest fixed period.

The delay distribution calculation unit 36 further accumulates the occurrence probabilities for each response time in order from the shorter response time, and thus the cumulative probability representing the probability that the actual response time is equal to or less than the response time for each response time. Is calculated.
The delay distribution calculation unit 36 stores the response time probability distribution and the cumulative probability distribution in the storage unit 32.
The delay distribution calculation unit 36 adds the response time stored between the previous calculation time and the current calculation time to the probability distribution of the response distribution calculated last time, and is used during the previous calculation. You may correct so that the response time which is not utilized at the time of calculation of may be excluded. By using the probability distribution obtained by such correction, the delay distribution calculation unit 36 can reduce the processing amount for obtaining the cumulative probability distribution.

ここでP(X)は、実際の応答時間が推定した応答時間X以内に収まる累積確率であり、記憶部３２に記憶された応答時間の累積確率分布を参照することにより求められる。またQ(X)は、その推定応答時間Xに応じた長さを持つメッセージを応答時間内に再生する場合のユーザの体感品質である。すなわち、（１）式の右辺の第１項は、応答時間の推定に成功したときのユーザの体感品質の期待値を表す。なお、メッセージ長は、推定応答時間Xと等しくてもよく、あるいは、推定応答時間Xから所定のオフセット時間（例えば、2秒〜4秒）を減じた長さであってもよい。また、(1-P(X))は、推定応答時間Xよりも実際の応答時間が長くなる確率を表す。そしてR(X)は、推定応答時間Xよりも実際の応答時間が長くなる場合に、「しばらくお待ち下さい」といった追加のメッセージを応答時間内に挿入したときのユーザの体感品質を表す。すなわち、（１）式の右辺の第２項は、応答時間の推定に失敗したときのユーザの体感品質の期待値を表す。なお、以下では、説明の便宜上、体感品質Q(X)を未延長体感品質と呼び、体感品質R(X)を延長体感品質と呼ぶ。 The message length determination unit 37 determines the length of the message voice reproduced during the response time based on the probability distribution of the response time so that the user's experience quality is as high as possible.
Therefore, the message length determination unit 37 maximizes the average experience quality S (X), which is an example of the evaluation value of the user experience quality corresponding to the estimated value X of the response time, calculated according to the following equation. Then, the estimated value X of the response time is determined.

Here, P (X) is a cumulative probability that falls within the estimated response time X, and is obtained by referring to the cumulative probability distribution of the response time stored in the storage unit 32. Q (X) is the quality of experience of the user when a message having a length corresponding to the estimated response time X is reproduced within the response time. That is, the first term on the right side of the equation (1) represents the expected value of the user's quality of experience when the response time is successfully estimated. The message length may be equal to the estimated response time X or may be a length obtained by subtracting a predetermined offset time (for example, 2 seconds to 4 seconds) from the estimated response time X. Further, (1-P (X)) represents the probability that the actual response time will be longer than the estimated response time X. R (X) represents the user's quality of experience when an additional message such as “Please wait for a while” is inserted within the response time when the actual response time is longer than the estimated response time X. That is, the second term on the right side of equation (1) represents the expected value of the user's experience quality when the response time estimation fails. Hereinafter, for convenience of explanation, the quality of experience Q (X) is referred to as an unextended experience quality, and the quality of experience R (X) is referred to as an extended experience quality.

  FIG. 4 is a diagram illustrating a change in response time of the user's unextended experience quality Q (X) when the response time is successfully estimated and the extended experience quality R (X) when the response time is unsuccessful. . In FIG. 4, the horizontal axis represents response time, and the vertical axis represents the user's quality of experience. The graph 401 represents the relationship between the response time X and the unextended experience quality Q (X), and the graph 402 represents the relationship between the response time X and the extended experience quality R (X). The unextended experience quality Q (X) and the extended experience quality R (X) are obtained from the results of subjective evaluation of the experience quality when the response time is variously changed by, for example, experiments in advance. . The quality of experience is expressed as, for example, a mean opinion score (MOS) value.

In general, the longer the response time, the lower the user experience quality. Therefore, as shown in the graph 401, the user's unextended quality of experience Q (X) when a message corresponding to the response time is reproduced decreases substantially monotonically as the response time increases. In general, when an additional message is inserted in addition to the original message within the response time, the quality of experience of the user is extremely lowered. In this case, the user's quality of experience is not greatly affected by the length of the response time, and is always low. Therefore, as shown in the graph 402, the user's extended experience quality R (X) is lower than the user's non-extended experience quality Q (X) and the response time is shorter, regardless of the response time. The difference between the unextended quality of experience Q (X) and the extended quality of experience R (X) becomes large.
Note that if the actual response time is longer than the estimated response time, and if no additional message is inserted within the response time, the silent period during which the system does not output any sound will be longer, so the user experience The quality is even worse than when additional messages are inserted.

  In order to obtain the user's unextended experience quality Q (X) and the extended experience quality R (X) according to the response time, for example, the user's unextended experience quality Q (X) and the extended experience experience for each sample point of the response time. A reference table representing the quality R (X) is stored in the storage unit 32 in advance. Then, the message length determination unit 37 may obtain the user's unextended experience quality Q (X) and extended experience quality R (X) corresponding to the response time X by referring to the reference table.

  The message length determination unit 37 calculates the cumulative probability P (X), the unextended experience quality Q (X), and the extended experience quality R (X) according to the estimated value X while changing the estimated value X of the response time. Thus, the average experience quality S (X) of the user corresponding to the estimated value X is calculated according to the equation (1). The message length determination unit 37 obtains an estimated value Xmax of response time (hereinafter, referred to as “maximum quality response time” for the sake of convenience) that maximizes the average experience quality S (X) of the user. Note that the message length determination unit 37 may calculate the estimated value Xmax using an optimal solution calculation method such as a gradient method. Then, the message length determination unit 37 determines the message length according to the maximum quality response time Xmax. The message length may be equal to the quality maximum response time Xmax, or the message length may be set to a value obtained by subtracting a predetermined offset time from the quality maximum response time Xmax as described above.

  The message length determination unit 37 notifies the message generation unit 38 of the determined message length.

  The message generation unit 38 generates text information of a message that is a source of message voice corresponding to the message length notified from the message length determination unit 37.

In the present embodiment, the message generator 38 creates message text information by combining a plurality of pre-registered standard messages and input keywords. The standard message is a message that follows the keyword such as “is” or “I will tell you” and is stored in the storage unit 32 in advance. In addition, the storage unit 32 stores a fixed message table representing the correspondence between a fixed message and the time length of the fixed message. Then, the message generation unit 38 sets the remaining time obtained by subtracting the length of the input keyword from the length of the message voice as the time length of the standard message. Then, the message generator 38 selects a fixed message having a time length closest to the set time length by referring to the fixed message table. The message generation unit 38 creates the text information of the message by combining the input keyword and the standard message, for example, by inserting the keyword at a specified position in the standard message. For example, if the input keyword is “news” and the selected standard message is “I will tell you”, the message generator 38 will use “I will tell you the news” as the text information of the message.
The message generator 38 notifies the text synthesizer 39 of the text information of the message. Further, when the time corresponding to the length of the message voice has elapsed after the control unit 33 receives the input voice signal, when the content is not ready to be transmitted to the terminal 2, the message generation unit 38 adds Is sent to the speech synthesizer 39.

  Note that according to the modification, the message generation unit 38 may use the fixed message itself having the determined message voice length as the text information of the message. In this case, the fixed message is a message having meaning only by the fixed message, for example, “Please wait for a while” or “Now searching”.

The voice synthesizer 39 creates a synthesized voice signal of text included in the content. The voice synthesizer 39 synthesizes the message voice based on the text information of the message reproduced during the response time.
The speech synthesizer 39 first converts the text to be synthesized into speech into phonetic information. The phonetic information is information representing the reading of the original text included in the text. For example, the phonetic information is information in which the reading of the original text is expressed by katakana characters, and further the position of the accent and the position of the break are added.

  The speech synthesis unit 39 reads a language dictionary stored in the storage unit 32 in order to convert text into phonetic information. In the language dictionary, for example, various words assumed to appear in the text information, readings of the words, parts of speech, and utilization forms are registered. For example, the speech synthesizer 39 performs morphological analysis on the original sentence included in the text using the language dictionary, and determines the reading of the text in the original sentence, the position of the accent, and the position of the break. At that time, for example, the speech synthesizer 39 sets a position where a punctuation mark is set in the original text as a break position.

  The speech synthesis unit 39 can use, for example, a method using dynamic programming as the morphological analysis. Then, the speech synthesizer 39 creates phonetic information according to the reading of each word, the accent position, and the break position.

  Next, the speech synthesizer 39 generates a target prosody for generating synthesized speech based on the phonetic information. For this purpose, the speech synthesis unit 39 reads a plurality of prosodic dictionaries from the storage unit 32. This prosodic dictionary stores prosodic models representing changes in voice pitch and phoneme length over time. Then, the speech synthesizer 39 applies the prosodic model that most closely matches the position in the sentence or the accent position indicated in the phonetic information from the prosodic dictionary. Then, the speech synthesizer 39 creates a target prosody corresponding to the phonetic information in accordance with the applied prosodic model and preset synthesis parameters. Note that the synthesis parameters include, for example, a parameter representing speech speed and a parameter representing voice pitch. Furthermore, the synthesis parameter may include a parameter representing inflection, volume, and the like. The target prosody includes the phoneme length and the pitch frequency for each phoneme that is a unit for determining a speech waveform. Further, the target prosody may include amplitude information of individual phoneme waveforms. Note that the phoneme can be, for example, one vowel or one consonant.

When the target prosody is determined, the speech synthesizer 39 creates a synthesized speech signal by, for example, a vocoder method or a waveform editing method.
For each phoneme, the speech synthesizer 39 selects a speech waveform closest to the phoneme length and pitch frequency of the target prosody from, for example, a plurality of speech waveforms registered in the speech waveform dictionary by pattern matching. For this purpose, the speech synthesizer 39 reads the speech waveform dictionary from the storage unit 32. The speech waveform dictionary records a plurality of speech waveforms and an identification number of each speech waveform. The voice waveform is, for example, a waveform signal extracted in units of phonemes from voice signals obtained by recording various voices in which one or more narrators read various texts.
Furthermore, the speech synthesizer 39 connects the selected speech waveform and the waveform pattern of the corresponding phoneme shown in the target prosody so that the speech waveform selected for each phoneme can be connected along the target prosody. The deviation amount may be calculated as waveform conversion information.
The speech synthesizer 39 creates waveform generation information including the identification number of the speech waveform selected for each phoneme. The waveform generation information may further include waveform conversion information.

The voice synthesizer 39 creates a synthesized voice signal based on the waveform generation information. For this purpose, the speech synthesizer 39 reads the speech waveform signal corresponding to the speech waveform identification number of each phoneme included in the waveform generation information from the speech waveform dictionary stored in the storage unit 32. Then, the speech synthesizer 39 creates a synthesized speech signal by connecting each speech waveform signal continuously. When the waveform conversion information is included in the waveform generation information, the speech synthesizer 39 continuously connects the speech waveform signals by correcting each speech waveform signal according to the waveform conversion information obtained for the corresponding phoneme. By doing so, a synthesized speech signal is created.
The control unit 33 outputs the message voice generated by the voice synthesis unit 39 to the terminal 2 via the communication unit 31 and the communication network 4. Thereafter, the control unit 33 outputs a synthesized voice signal of text included in the content to the terminal 2 via the communication unit 31 and the communication network 4.
If the control unit 33 is not ready to transmit the synthesized audio signal of the content to the terminal 2 even after the time corresponding to the length of the message audio has elapsed after receiving the input audio signal, the audio signal of the additional message Is output to the terminal 2 via the communication unit 31 and the communication network 4.

FIG. 5 is an operation sequence diagram of dialogue control processing in the voice dialogue system 1.
First, the processing unit 25 of the terminal 2 acquires an input voice signal corresponding to the voice uttered by the user via the voice input unit 21 (step S101). Then, the processing unit 25 transmits the input audio signal to the server 3 via the communication unit 23 and the communication network 4. In addition, the processing unit 25 starts measuring the elapsed time from the time when the input audio signal ends.

  When the server 3 receives the input voice signal, the voice recognition unit 34 of the control unit 33 extracts keywords included in the input voice signal (step S102).

  Then, the delay distribution calculation unit 36 of the control unit 33 obtains the cumulative probability distribution of the response time based on the response time history of the latest fixed period (step S103). Then, the delay distribution calculation unit 36 stores the cumulative probability distribution of response times in the storage unit 32.

The message length determination unit 37 of the control unit 33 determines the message length so that the average experience quality, which is the sum of the expected value of the unexpanded experience quality and the expected value of the extended experience quality, is maximized (step S104).
The message generation unit 38 of the control unit 33 obtains text information of a message that is reproduced during the response time according to the determined message length (step S105).

  The voice synthesis unit 39 of the control unit 33 synthesizes the voice of the message based on the text information of the message (step S106). Then, the server 3 outputs the synthesized message voice to the terminal 2. Further, the server 3 synthesizes the voice of the additional message as necessary, and outputs the synthesized additional message voice to the terminal 2.

  The processing unit 25 of the terminal 2 causes the voice output unit 22 to reproduce the message voice (step S107). Further, the processing unit 25 reproduces the voice of the additional message when the synthesized voice signal of the content is not received even after the reproduction of the message voice is finished. Note that the processing unit 25 may repeatedly reproduce the voice of the additional message until the synthesized voice signal of the content is received.

  On the other hand, the search unit 35 of the control unit 33 specifies the address information of the information source corresponding to the keyword (step S108). Then, the search unit 35 acquires content from the information source indicated by the address information (step S109).

The voice synthesizer 39 synthesizes a voice signal corresponding to the text included in the content (step S110). Then, the server 3 notifies the terminal 2 of the synthesized speech signal of the content and the text included in the content.
When receiving the content and the synthesized speech signal of the text included in the content, the processing unit 25 of the terminal 2 causes the speech output unit 22 to reproduce the synthesized speech signal (step S111).
Then, the voice dialogue system 1 ends the dialogue control process.

  As described above, this voice interactive apparatus is configured to determine the length of the message voice reproduced during the response time when the message is additionally inserted when the probability distribution of the response time and the estimation of the response time fail. Considering the experience quality, the decision is made so that the experience quality is the highest. Therefore, even if the response time varies, this voice interaction device can set the length of the message voice so that the user's quality of experience is statistically best.

  When the communication time between the server 3 and the terminal 2 can be regarded as substantially constant, the control unit 33 of the server 3 waits until the synthesized voice of the content is ready for transmission after the audio signal received from the terminal 2 is completed. You may keep time. Then, the control unit 33 may obtain a value obtained by adding twice the communication time between the terminal 2 and the server 3 to the measured time as a response time, and store the response time in the storage unit 32.

According to the modification, the message length determination unit 37 may determine the message length according to the following procedure in order to reduce the amount of calculation for determining the message length.
In this example, the message length determination unit 37 firstly includes relational expressions Q (P), R representing the relationship between the cumulative probability of response time and the user's unextended experience quality Q (X), extended experience quality R (X). Calculate (P).

  FIG. 6A is a diagram illustrating an example of a function representing the relationship between response time and bodily sensation quality. FIG. 6B is a diagram illustrating an example of the cumulative probability of response time. FIG. 6C is a diagram illustrating the relationship between the cumulative probability of response time and the user's quality of sensation obtained from FIGS. 6A and 6B. In FIG. 6A and FIG. 6B, the horizontal axis represents response time. In FIG. 6A, the vertical axis represents the sensation quality value, and in FIG. 6B, the vertical axis represents the cumulative probability. A graph 601 represents the relationship between the unextended experience quality and the response time. A graph 602 represents the relationship between the extended experience quality and the response time. A graph 603 represents the cumulative probability of response time. In FIG. 6C, the horizontal axis represents the cumulative probability, and the vertical axis represents the quality of experience. The graph 604 represents the relationship between the unextended experience quality and the cumulative probability of response time. A graph 605 represents the relationship between the extended experience quality and the cumulative probability of response time.

  For each response time, the message length determination unit 37 refers to the graphs 601 and 602 to obtain an unextended experience quality value Q (X) and an extended experience quality R (X) corresponding to the response time X. Further, the message length determination unit 37 refers to the graph 603 for each response time, and obtains the cumulative probability P (X) corresponding to the response time X. Then, the message length determination unit 37 obtains the graphs 604 and 605 by associating the cumulative probability P (X) corresponding to the same response time with the unextended experience quality value Q (X) and the extended experience quality R (X). .

なお、S(P)は、応答時間の累積確率Pの関数として表されたユーザの平均体感品質であり、Q(P)は、累積確率Pの関数として表された、未延長体感品質を表す。またR(P)は、累積確率Pの関数として表された延長体感品質を表す。
（２）式及び図６（ｃ）を参照すると、ユーザの平均体感品質S(P)は、累積確率Pを底辺、未延長体感品質Q(P)を高さとする長方形６１１の面積と、(1-P)を底辺、延長体感品質R(P)を高さとする長方形６１２の面積の和で表される。したがって、メッセージ長決定部３７は、この二つの長方形６１１、６１２の和が最大となる累積確率Pを決定する。そしてメッセージ長決定部３７は、累積確率Pと応答時間の関係を表す参照テーブルを参照することにより、その累積確率Pに対応する応答時間を求めればよい。 Here, since the response time X and the cumulative probability P have a one-to-one correspondence, the average experience quality S (X) of the user expressed by the formula (1) is the cumulative response time as shown in the following formula. It can also be expressed as a function of probability P.

Note that S (P) is the average experience quality of the user expressed as a function of the cumulative probability P of response time, and Q (P) represents the unextended experience quality expressed as a function of the cumulative probability P. . R (P) represents the extended experience quality expressed as a function of the cumulative probability P.
Referring to the equation (2) and FIG. 6C, the average experience quality S (P) of the user is the area of the rectangle 611 having the cumulative probability P as the base and the unextended experience quality Q (P) as the height. It is expressed as the sum of the areas of a rectangle 612 with 1-P) as the bottom and the extended quality of experience R (P) as the height. Therefore, the message length determination unit 37 determines the cumulative probability P that maximizes the sum of the two rectangles 611 and 612. Then, the message length determination unit 37 may obtain a response time corresponding to the cumulative probability P by referring to a reference table representing the relationship between the cumulative probability P and the response time.

Furthermore, in order to reduce the amount of calculation for obtaining the cumulative probability P that maximizes the sum of the two rectangles, the message length determination unit 37 may perform approximation described below.
As described above, since the extended quality of experience is not significantly affected by the response time, it is not significantly affected by the cumulative probability of the response time. Therefore, the function representing the relationship between the extended experience quality and the cumulative probability P can be approximated as a linear function having a constant experience quality value R.
In addition, the user's unextended quality of experience Q (X) when a message having a message length determined according to the estimated response time is reproduced within the response time monotonously decreases with respect to the response time. As the probability P increases, the unextended quality of experience Q (P) decreases monotonously. Therefore, the relationship between the cumulative probability P and the unextended experience quality Q (P) is expressed by linear approximation.

  FIG. 7 is a graph showing the relationship between the cumulative probability P and the bodily sensation quality shown in FIG. 6C, where the bodily sensation quality R (P) is a constant value R (0), and the cumulative probability P is not extended. A graph that approximates the relationship of the quality of experience Q (P) with a straight line is shown. The horizontal axis represents the cumulative probability, and the vertical axis represents the quality of experience. A straight line 701 represents the relationship between the cumulative probability P approximated by a straight line and the unextended experience quality Q (P). A straight line 702 represents the extended experience quality R (P) approximated as a constant value R. The average quality of experience S (P) of the user is represented by a rectangle 711, which is a product of cumulative probabilities P and Q (P), and a product of (1-P) and R (0), which is represented by a rectangle 712. The sum of Here, the rectangle 711 is divided into a rectangle 711a and a rectangle 711b by a line of the sensation quality value R. The sum of the area P × R of the rectangle 711a and the area (1-P) × R of the rectangle 712 is R regardless of the cumulative probability P. Therefore, in order to maximize the average quality of experience S (P), the message length determination unit 37 may determine the cumulative probability P so that the area of the rectangle 711b is maximized. Here, the rectangle 711b is inscribed in a right triangle having points A, B, and C as vertices. However, the points A and B are intercepts of the straight lines 701 and 702, respectively, and the point C is an intersection of the straight lines 701 and 702. Since the area of the rectangle inscribed in the right triangle is known to be the case where the length of the bottom of the rectangle is 1/2 of the length of the right triangle, the area of the rectangle 711b is the cumulative probability P Is the maximum when the cumulative probability Pc at the intersection C is ½ of the value. Therefore, the message length determination unit 37 may calculate the cumulative probability Pc at the intersection C in order to determine the message length. Then, if the cumulative probability (Pc / 2) is 1 or less, the message length determination unit 37 refers to the reference table that represents the relationship between the response time and the cumulative probability, thereby responding to the cumulative probability (Pc / 2). What is necessary is just to obtain | require time and to determine message length according to the response time. On the other hand, if the cumulative probability (Pc / 2) is greater than 1, the message length determination unit 37 refers to the reference table representing the relationship between the response time and the cumulative probability, so that the shortest response time at which the cumulative probability is one And the message length may be determined according to the response time.

  According to this method, the message length determination unit 37 determines only the response time corresponding to the cumulative probability (Pc / 2) without changing the response time by trial and error, and thus determines the message length. The amount of computation required at the time can be greatly reduced.

  Note that the message length determination unit 37 determines the difference (Q (0) between the unextended experience quality Q (0) and the extended experience quality R (0) at the minimum response time when the cumulative probability P (X) is initially greater than 0. ) −R (0)) divided by the slope a of the straight line 701 ((Q (0) −R (0)) / a) may be multiplied by 1/2 to calculate the cumulative probability. The cumulative probability obtained as a result of this calculation is also the same value as the cumulative probability (Pc / 2).

ここで、αは、応答時間の確率分布における、最大確率であり、X_αは、その最大確率αに相当する応答時間である。またP_αは、応答時間X_αの累積確率である。 In addition, the message length determination unit 37, instead of calculating the cumulative probability Pc corresponding to the intersection C every time the probability distribution of the response time is obtained, the unextended experience quality Q (X) and the extended experience quality R (X) You may obtain | require the response time Xc from which becomes equal. The response time Xc is obtained from the relationship between the response time X, the user's unextended experience quality Q (X), and the extended experience quality R (X). Then, the message length determination unit 37 may obtain the cumulative probability Pc based on the response time Xc. In this case, if Xc is equal to or shorter than the shortest response time Xm at which the cumulative probability is 1, the message length determination unit 37 refers to the reference table representing the relationship between the response time and the cumulative probability, so that the cumulative corresponding to Xc The probability Pc may be obtained. On the other hand, if Xc is larger than the response time Xm, the message length determination unit 37 calculates the cumulative probability Pc approximately corresponding to Xc, for example, according to the following equation.

Here, α is a maximum probability in the probability distribution of response time, and X _α is a response time corresponding to the maximum probability α. P _α is the cumulative probability of the response time X _α .

FIG. 8 is a diagram illustrating the relationship between the parameters used in equation (3). The upper graph in FIG. 8 is a graph representing the relationship between the response time and the cumulative probability, and the lower graph in FIG. 8 is a graph representing the distribution of the occurrence probability for each response time. (3) alpha in the equation, the inflection point of the cumulative probability distribution corresponds to the slope of the tangent 801 of the accumulated probability distribution curve in response time X _alpha.

  Since Xc can be calculated regardless of the probability distribution of response time, for example, it may be calculated in advance and stored in the storage unit 32. Therefore, in this modification, the message length determination unit 37 can obtain the response time that maximizes the average experience quality without obtaining the relational expression between the cumulative probability and the experience quality, so that the amount of calculation can be further reduced. .

In general, the shorter the response time, the higher the quality of the user's experience, so when the actual response time is shorter than the estimated response time, the voice interactive device interrupts the message halfway and It is preferable to speed up regeneration as much as possible.
Therefore, according to another modified example, the message length determination unit 37 sets the message length so that the probability of being less than the suspendable position in the message and the response time is high.

  The position where the message can be interrupted is preferably a position where even if the message is interrupted at the interruptable position, the user can naturally hear the end of the message at the interrupt position. For example, when a message is formed by the combination of the input keyword "news" and the standard message "we will tell you about", between "news" and "about" or between "about" and "let me tell" The message can be interrupted.

  In the above example, if the actual response time ends before “we tell you”, the control unit 33 interrupts the message voice at “about” and the content voice terminal 25. You can start output to. Therefore, the voice interaction apparatus can be expected to improve the quality of the user's experience due to the shortened response time. On the other hand, when the actual response time ends in the middle of “I will tell you”, the control unit 33 outputs the audio of the content to the terminal 2 after finishing outputting the message audio to the end. Even if the time is shortened, the quality of experience of the user is not improved.

  Therefore, the message length determination unit 37 sets a position where the fixed message can be interrupted so that the probability that the message can be interrupted is increased as much as possible when the actual response time is shorter than the estimated response time.

  In this modified example, the fixed message table stored in the storage unit 32 includes, for each fixed message, a breakable position, a time from the start of the fixed message to the breakable position, and two consecutive breakable positions. The time is stored.

ここで、X_i(i=1,2,..,n-1)は、中断可能位置に相当する時間である（なお、メッセージが入力キーワード＋定型メッセージで形成される場合、X₁は、入力キーワードの時間長に相当する）。なお、応答時間中に再生されるメッセージが定型メッセージのみで形成される場合には、X₁は、定型メッセージの先頭から最初の中断可能位置までの時間長となる。X_nは、メッセージ全体の長さである。またP(X_i)は、応答時間開始からの時間X₁の累積確率である。Q(X_i)は、応答時間がX_iである場合のユーザの未延長体感品質である。そしてR(X)は、推定された応答時間がXである場合における、メッセージが追加挿入された時のユーザの延長体感品質である。
メッセージ長決定部３７は、定型メッセージテーブルに格納された各定型メッセージの中断可能位置の組み合わせ{X₁,X₂,…,X_n}の中から、平均体感品質Eが最大となる組み合わせ{X₁,X₂,…,X_n}を算出すればよい。そして定型メッセージ生成部３８は、その組み合わせ{X₁,X₂,…,X_n}からの誤差の二乗和が最小となる定型メッセージを選択を利用して、メッセージを生成すればよい。 The message length determination unit 37 calculates the average experience quality E of the user according to the following equation for each fixed message in which the total length of the input keyword and the fixed message is equal to or less than the maximum response time.

Here, X _i (i = 1, 2, .., n−1) is a time corresponding to the position where interruption is possible (in the case where the message is formed by the input keyword + standard message, X ₁ is Equivalent to the time length of the input keyword). In the case where the message to be played during the response time is formed only in fixed message, X ₁ is a time length from the beginning of the fixed message to the first interruptible position. X _n is the length of the entire message. P (X _i ) is the cumulative probability of time X ₁ from the start of the response time. Q (X _i ) is the user's unextended quality of experience when the response time is X _i . R (X) is the extended experience quality of the user when a message is additionally inserted when the estimated response time is X.
The message length determination unit 37 selects a combination {X ₁ , X ₂ ,..., X _n } having the maximum average experience quality E from the combinations {X ₁ , X ₂ ,..., X _n } of the fixed messages stored in the fixed message table. ₁ , X ₂ ,..., X _n } may be calculated. Then, the fixed message generator 38 may generate a message by using selection of a fixed message that minimizes the sum of squares of errors from the combination {X ₁ , X ₂ ,..., X _n }.

Note that the message length determination unit 37 may determine a message length based on a graph representing the relationship between the cumulative probability of response time and the user's experience quality, as in the above-described modification. In this case, as in the above-described modification, each term of equation (4) has the cumulative probability difference (P _i -P _i-1 ) as the base and the unextended experience quality Q (P _i ) as the height. It corresponds to the rectangular area. Therefore, the message length determination unit 37 can calculate the average experience quality E as the sum of the areas of the rectangles, so that it is possible to reduce the amount of calculation when determining the message length that maximizes the average experience quality E.

  When transmitting the message voice to the terminal 2, the control unit 33 also transmits information indicating the interruptible position to the terminal 2 together with the message voice. Thus, even when the terminal 2 receives the synthesized voice signal of the content from the server 3 during the playback of the message voice, the terminal 2 can stop the playback of the message voice at the nearest interruptible position and start the playback of the synthesized voice signal. .

  According to this modified example, when the actual response time is shorter than the estimated response time, the voice interactive device can interrupt the message halfway and increase the probability that the user's quality of experience can be improved.

  According to still another modification, the processing unit 25 of the terminal 2 may have functions of a speech recognition unit and a speech synthesis unit. Thereby, the processing load of the server 3 is reduced.

  According to still another modification, each unit included in the terminal and each unit included in the server in the above-described embodiment or the modification thereof may be mounted on one device. One or a plurality of processors included in the apparatus may execute each function of the processing unit of the terminal and each function of the control unit of the server.

  Furthermore, a computer program that causes a computer to realize each function of the control unit of the server according to each of the above embodiments or modifications thereof may be provided in a form recorded on a computer-readable recording medium. The computer-readable recording medium can be, for example, a magnetic recording medium, an optical recording medium, or a semiconductor memory. However, the recording medium does not include a carrier wave.

  All examples and specific terms listed herein are intended for instructional purposes to help the reader understand the concepts contributed by the inventor to the present invention and the promotion of the technology. It should be construed that it is not limited to the construction of any example herein, such specific examples and conditions, with respect to showing the superiority and inferiority of the present invention. Although embodiments of the present invention have been described in detail, it should be understood that various changes, substitutions and modifications can be made thereto without departing from the spirit and scope of the present invention.

The following supplementary notes are further disclosed regarding the embodiment described above and its modifications.
(Appendix 1)
A voice recognition unit that extracts a predetermined keyword from a voice signal representing the user's voice collected by the voice input unit;
A search unit for searching for content according to the keyword;
A delay distribution calculating unit that calculates a probability distribution of response time from the end of input of the user's voice to presentation of the content to the user;
Based on the probability distribution, when the response time is equal to or less than the estimated response time, the estimated response time or a first message voice having a length obtained by subtracting a predetermined offset from the estimated response time is included in the response time. The first expected value of the first experience quality of the user when playing back, the response time longer than the estimated response time, and the first message voice and the additional second message within the response time A message length determination unit that determines the length of the message voice so that the evaluation value that is the sum of the second expected values of the second experience quality of the user when the voice is reproduced is maximized;
A message generator for generating text information of the first message having a length of the message voice;
A speech synthesizer that synthesizes the first message speech from text information of the first message;
Spoken dialogue apparatus having
(Appendix 2)
The message length determination unit may determine the first expected value based on the probability distribution, the response time being shorter than the estimated response time, and at least one of the interruptable positions of the first message voice. Calculated as the sum of the expected value of the user's bodily sensation quality when playback is interrupted at an interruptible position below the response time and the expected value when the entire first message voice is played back within the response time The voice interactive apparatus according to appendix 1.
(Appendix 3)
A storage unit for storing text information about each of the plurality of fixed messages and the interruptible position;
The message generation unit selects a fixed message having a suspendable position closest to the suspendable position when the evaluation value is maximum from the plurality of fixed messages, and includes the selected fixed message. The spoken dialogue apparatus according to claim 2, wherein the text information of the first message is generated.
(Appendix 4)
The message length determination unit, from the relationship between the response time and the first experience quality, and the cumulative probability of the response time in which the occurrence probability of the response time is accumulated in order from the shortest response time, the cumulative probability and the A relational expression obtained by linearly approximating the relationship with the first bodily sensation quality is obtained, and a response time corresponding to half of the cumulative probability when the value of the first bodily sensation quality is equal to the second bodily sensation quality in the relational expression. Alternatively, the voice interactive device according to appendix 1, wherein the length of the first message voice is set at a time obtained by subtracting the predetermined offset from the response time.
(Appendix 5)
The message length determination unit, based on the relationship between the response time and the first and second quality of experience, and the cumulative probability of the response time obtained by accumulating the probability of occurrence of the response time in order from the shortest response time A first relational expression representing the relationship between the probability and the first bodily sensation quality and a second relational expression representing the relationship between the cumulative probability and the first bodily sensation quality are obtained and correspond to the estimated response time. A product of a first cumulative probability and the first experience quality corresponding to the first cumulative probability determined according to the first relational expression is calculated as the first experience quality, and the second The product of the second sensory quality according to the first cumulative probability obtained according to the relational expression and a value obtained by subtracting the first cumulative probability from 1 is calculated as the second sensory quality. The voice interactive apparatus according to 1.
(Appendix 6)
A storage unit that stores the response time each time the audio signal is obtained from the audio input unit;
The spoken dialogue apparatus according to appendix 1, wherein the delay distribution calculation unit calculates the probability distribution from a plurality of frequency distributions of the response times stored within a predetermined period in the past.
(Appendix 7)
A predetermined keyword is extracted from the voice signal representing the voice of the user collected by the voice input unit;
Search for content according to the keyword,
Calculating the probability distribution of response time from the end of the user's voice input to the presentation of the content to the user;
Based on the probability distribution, when the response time is equal to or less than the estimated response time, the estimated response time or a first message voice having a length obtained by subtracting a predetermined offset from the estimated response time is included in the response time. The first expected value of the first experience quality of the user when playing back, the response time longer than the estimated response time, and the first message voice and the additional second message within the response time The length of the message voice is determined so that the evaluation value that is the sum of the second expected values of the second experience quality of the user when the voice is reproduced is maximized,
Generating text information of the first message having the length of the message voice;
Synthesizing the first message voice from text information of the first message;
Dialog control method including that.

1 Spoken Dialogue System 2 Terminal 3 Server (Speech Dialogue Device)
DESCRIPTION OF SYMBOLS 4 Communication network 5 External information source 21 Voice input part 22 Voice output part 23 Communication part 24 Storage part 25 Processing part 31 Communication part 32 Storage part 33 Control part 34 Voice recognition part 35 Search part 36 Delay distribution calculation part 37 Message length determination part 38 Message generator 39 Speech synthesizer

Claims (5)

A voice recognition unit that extracts a predetermined keyword from a voice signal representing the user's voice collected by the voice input unit;
A search unit for searching for content according to the keyword;
A delay distribution calculating unit that calculates a probability distribution of response time from the end of input of the user's voice to presentation of the content to the user;
Based on the probability distribution, when the response time is equal to or less than the estimated response time, the estimated response time or a first message voice having a length obtained by subtracting a predetermined offset from the estimated response time is included in the response time. The first expected value of the first experience quality of the user when playing back, the response time longer than the estimated response time, and the first message voice and the additional second message within the response time A message length determination unit that determines the length of the message voice so that the evaluation value that is the sum of the second expected values of the second experience quality of the user when the voice is reproduced is maximized;
A message generator for generating text information of the first message having a length of the message voice;
A speech synthesizer that synthesizes the first message speech from text information of the first message;
Spoken dialogue apparatus having   The message length determination unit may determine the first expected value based on the probability distribution, the response time being shorter than the estimated response time, and at least one of the interruptable positions of the first message voice. Calculated as the sum of the expected value of the user's bodily sensation quality when playback is interrupted at an interruptible position below the response time and the expected value when the entire first message voice is played back within the response time The voice interactive apparatus according to claim 1. A storage unit for storing text information about each of the plurality of fixed messages and the interruptible position;
The message generation unit selects a fixed message having a suspendable position closest to the suspendable position when the evaluation value is maximum from the plurality of fixed messages, and includes the selected fixed message. The spoken dialogue apparatus according to claim 2, wherein text information of the first message is generated.   The message length determination unit, from the relationship between the response time and the first experience quality, and the cumulative probability of the response time in which the occurrence probability of the response time is accumulated in order from the shortest response time, the cumulative probability and the A relational expression obtained by linearly approximating the relationship with the first bodily sensation quality is obtained, and a response time corresponding to half of the cumulative probability when the value of the first bodily sensation quality is equal to the second bodily sensation quality in the relational expression. The voice interactive device according to claim 1, wherein the length of the first message voice is set at a time obtained by subtracting the predetermined offset from the response time. A predetermined keyword is extracted from the voice signal representing the voice of the user collected by the voice input unit;
Search for content according to the keyword,
Calculating the probability distribution of response time from the end of the user's voice input to the presentation of the content to the user;
Based on the probability distribution, when the response time is equal to or less than the estimated response time, the estimated response time or a first message voice having a length obtained by subtracting a predetermined offset from the estimated response time is included in the response time. The first expected value of the first experience quality of the user when playing back, the response time longer than the estimated response time, and the first message voice and the additional second message within the response time The length of the message voice is determined so that the evaluation value that is the sum of the second expected values of the second experience quality of the user when the voice is reproduced is maximized,
Generating text information of the first message having the length of the message voice;
Synthesizing the first message voice from text information of the first message;
Dialog control method including that.

Images