JP2017106989A - Voice interactive device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice interactive device and a program that achieve natural voice interactions.SOLUTION: A voice interactive device 100C comprises: a voice input device 24 for collecting utterance voices Vx to generate utterance signals X; a voice acquisition part 32 for acquiring the utterance signals X; and a response generation part 36C that analyzes the utterance signals X and allows a reproducer 26 to selectively reproduce first response voices Vy1 for representing responses in response to the utterance voices Vx and second response voices Vy2 other than the responses.SELECTED DRAWING: Figure 10

Description

  The present invention relates to a voice dialogue technique for reproducing a response voice to a spoken voice.

  2. Description of the Related Art Conventionally, a voice dialogue technique has been proposed that realizes a dialogue with a user by reproducing a voice of a response to an utterance by a user (for example, an answer to a question). For example, Patent Document 1 discloses a technique for analyzing utterance contents by voice recognition on a user's uttered voice and synthesizing and reproducing a response voice according to the analysis result.

JP 2012-128440 A

  However, under the existing technologies including Patent Document 1, it is actually difficult to realize a natural voice dialogue that faithfully reflects the tendency of dialogue between real people, which is mechanical and unnatural. There is a problem that the user can take an impression. In view of the above circumstances, an object of the present invention is to realize a natural voice conversation.

  In order to solve the above-described problems, a voice interaction apparatus according to a preferred aspect of the present invention includes a voice acquisition unit that acquires an utterance signal that represents an utterance voice, a first response voice that represents a response to the utterance voice, A response generation unit that selectively causes the playback device to play back the second response sound other than the response. In the above-described aspect, the first response sound that indicates the answer to the uttered voice and the second response sound other than the answer are selectively played back from the playback device. Therefore, it is possible to realize a natural voice dialogue that simulates the tendency of an actual dialogue in which not only the talker's utterance but also the questioning / replying (rehearsal) to the talker occurs appropriately.

  The spoken dialogue apparatus according to a preferred aspect of the present invention includes a speech analysis unit that identifies a prosodic index value representing a prosody of an uttered speech from an utterance signal, and the response generation unit obtains a prosodic index value and a threshold value of the uttered speech. Comparison is made, and either the first response sound or the second response sound is selected according to the comparison result. In the above aspect, since either the first response speech or the second response speech is selected according to the result of comparing the prosodic index value representing the prosody of the speech speech with the threshold value, the prosody of the speech speech changes. It is possible to realize a natural voice conversation that simulates the tendency of an actual conversation that it is difficult to listen to the spoken voice and the possibility of answering questions increases.

  In a preferred aspect of the present invention, the speech analysis unit sets a representative value of prosodic index values in a plurality of past speech speech as a threshold value. In the above aspect, since the representative value of the prosodic index value in a plurality of past utterances is set as the threshold, the prosody of the utterance of the utterer is the standard prosody of the utterer (that is, the prosody assumed by the conversation partner). It is possible to realize a natural voice conversation that simulates the tendency of an actual conversation that a question is easily returned from the conversation partner when it deviates from).

  In a preferred aspect of the present invention, the response generation unit selects the first response speech when the prosodic index value is a numerical value outside the predetermined range including the threshold value, and when the prosodic index value is a numerical value inside the predetermined range, 2 Select response voice. In the above aspect, the first response voice is selected when the prosodic index value is a numerical value outside the predetermined range, and the second response voice is selected when the numerical value is inside the predetermined range. It is possible to reduce the possibility that the first response sound is reproduced with frequency.

  In a preferred aspect of the present invention, the response generation unit reproduces the first response voice for the utterance voice randomly selected from the plurality of utterance voices. In the above aspect, since the first response voice is reproduced with respect to the utterance voice randomly selected from the plurality of utterance voices, the tendency of the actual voice conversation in which the answer to the utterance voice is randomly generated is simulated. Natural speech dialogue can be realized. For example, the response generation unit variably sets the reproduction frequency of the first response sound for a plurality of speech sounds. It is also possible to set the reproduction frequency of the first response voice according to the use history of the voice conversation.

It is a block diagram of the voice interactive apparatus of 1st Embodiment. It is a flowchart of operation | movement of the voice interactive apparatus in 1st Embodiment. It is explanatory drawing of the speech sound and response sound in 1st Embodiment. It is explanatory drawing of the speech sound and response sound in 1st Embodiment. It is a flowchart of the response generation process of 1st Embodiment. It is a block diagram of the voice interactive apparatus of 2nd Embodiment. It is explanatory drawing of the speech voice and response voice in 2nd Embodiment. It is explanatory drawing of the speech voice and response voice in 2nd Embodiment. It is a flowchart of the response production | generation process in 2nd Embodiment. It is a block diagram of the voice interactive apparatus of 3rd Embodiment. It is a flowchart of operation | movement of the voice interactive apparatus in 3rd Embodiment. It is a flowchart of the response generation process in 3rd Embodiment. It is explanatory drawing of the speech sound and response sound in 3rd Embodiment. It is explanatory drawing of the speech sound and response sound in 3rd Embodiment. It is a block diagram of the voice interactive apparatus of 4th Embodiment. It is a flowchart of operation | movement of the voice interactive apparatus in 4th Embodiment. It is a flowchart of the response generation process in 4th Embodiment. It is explanatory drawing of the speech sound and response sound in 4th Embodiment. It is explanatory drawing of the speech sound and response sound in 4th Embodiment.

<First Embodiment>
FIG. 1 is a configuration diagram of a voice interactive apparatus 100A according to the first embodiment of the present invention. The voice interaction apparatus 100A of the first embodiment is a voice interaction system that reproduces a response voice (hereinafter referred to as “response voice”) Vy to a voice (hereinafter referred to as “utterance voice”) Vx produced by the user U. For example, a portable information processing device such as a mobile phone or a smartphone, or an information processing device such as a personal computer can be used as the voice interactive device 100A. It is also possible to realize the voice interaction device 100A in the form of a toy (for example, a doll such as a stuffed animal) or a robot simulating the appearance of an animal or the like.

  The utterance voice Vx is an utterance voice including, for example, a question (question) and a talk, and the response voice Vy is a reply voice including an answer to the question and an answer to the talk. The response voice Vy includes a voice meaning an interjection, for example. Interjections are independent words (impression verbs and exclamations) that are used independently of other segments and are not used. Specifically, phrases such as “Ye” and “Ee” that indicate the confession of the utterance, words such as “Eto” and “Ano ~” that indicate speech (stagnation of response), and responses (affirmation / "Yes", "No", etc., which expresses (No), words such as "Oh", "O", etc., which express the impression of the speaker, or "E?" A phrase such as “?” Can be exemplified as an interjection.

  The spoken dialogue apparatus 100A of the first embodiment generates a prosodic response voice Vy corresponding to the prosody of the uttered voice Vx. Prosodic is a linguistic and phonetic characteristic that can be perceived by the listener of a speech, and cannot be grasped only from the general notation of the language (for example, not a special notation that expresses prosody). Means nature. Prosody can be rephrased as a characteristic that allows the listener to recall or guess the intention or emotion of the speaker. Specifically, intonation (changes in tone of the voice, intonation), tone (speech level and strength), tone length (speech length), speech speed, rhythm (structure of temporal change in tone), accent (high and low) Various features, such as strong or weak accents, can be included in the concept of prosody, but typical examples of prosody are pitch (fundamental frequency) or volume.

  As illustrated in FIG. 1, the voice interaction apparatus 100A of the first embodiment includes a control device 20, a storage device 22, a voice input device 24, and a playback device 26. The voice input device 24 is an element that generates, for example, a voice signal (hereinafter referred to as “speech signal”) X representing the utterance voice Vx of the user U, and includes a sound collection device 242 and an A / D converter 244. The sound collection device (microphone) 242 collects the utterance voice Vx produced by the user U and generates an analog voice signal representing the sound pressure fluctuation of the utterance voice Vx. The A / D converter 244 converts the audio signal generated by the sound collection device 242 into a digital speech signal X.

  The control device 20 is an arithmetic processing device (for example, a CPU) that controls each element of the voice interactive device 100A. The control device 20 of the first embodiment acquires the utterance signal X supplied from the voice input device 24, and generates a response signal Y representing the response voice Vy with respect to the utterance voice Vx. The playback device 26 is an element that plays back the response sound Vy corresponding to the response signal Y generated by the control device 20, and includes a D / A converter 262 and a sound emitting device 264. The D / A converter 262 converts the digital response signal Y generated by the control device 20 into an analog audio signal, and the sound emitting device 264 (for example, a speaker or headphones) responds according to the converted audio signal. Vy is emitted as sound waves. The playback device 26 may include a processing circuit such as an amplifier that amplifies the response signal Y.

  The storage device 22 stores programs executed by the control device 20 and various data used by the control device 20. For example, a known recording medium such as a semiconductor recording medium or a magnetic recording medium, or a combination of a plurality of recording media can be arbitrarily employed as the storage device 22. The memory | storage device 22 of 1st Embodiment memorize | stores the audio | voice signal Z showing the response audio | voice of specific utterance content. In the following description, a case where a voice signal Z of a response voice such as “Yes”, which means a match, which is an example of an interjection, is stored in the storage device 22 will be exemplified. The audio signal Z is recorded in advance and stored in the storage device 22 as an audio file of an arbitrary format such as wav format.

  The control device 20 executes a program stored in the storage device 22 and thereby has a plurality of functions (voice acquisition unit 32, voice analysis unit 34A, response generation unit 36A) for establishing a dialogue with the user U. Realize. A configuration in which the function of the control device 20 is realized by a plurality of devices (that is, a system) or a configuration in which a dedicated electronic circuit shares a part of the function of the control device 20 may be employed.

  The voice acquisition unit 32 in FIG. 1 acquires an utterance signal X representing the utterance voice Vx. The voice acquisition unit 32 of the first embodiment acquires the speech signal X generated by the voice input device 24 from the voice input device 24. The voice analysis unit 34A specifies the pitch (fundamental frequency) P of the uttered voice Vx from the utterance signal X acquired by the voice acquisition unit 32. The pitch P is specified sequentially in a predetermined cycle. That is, the pitch P is specified for each of a plurality of different time points on the time axis. A known technique can be arbitrarily adopted to specify the pitch P of the speech voice Vx. Note that the pitch P can be specified by extracting an acoustic component in a specific frequency band from the speech signal X. The frequency band to be analyzed by the voice analysis unit 34A is variably set according to, for example, an instruction from the user U (for example, designation of male voice / female voice). It is also possible to dynamically change the frequency band to be analyzed according to the pitch P of the speech voice Vx.

  The response generation unit 36A causes the playback device 26 to play back the response voice Vy for the utterance voice Vx of the utterance signal X acquired by the voice acquisition unit 32. Specifically, the response generation unit 36A generates a response signal Y of the response voice Vy triggered by the pronunciation of the uttered voice Vx by the user U, and supplies the response signal Y to the playback device 26, thereby responding to the response voice Vy. Is played back by the playback device 26. The response generation unit 36A of the first embodiment adjusts the prosody of the audio signal Z stored in the storage device 22 in accordance with the pitch P of the uttered voice Vx specified by the audio analysis unit 34A, so that the response voice Vy The response signal Y is generated. That is, the response voice Vy obtained by adjusting the initial response voice represented by the voice signal Z in accordance with the prosody of the uttered voice Vx is played from the playback device 26.

  In an actual human-to-human conversation, the conversation partner speaks the response voice for the utterance voice with a pitch corresponding to the pitch near the end point of the utterance voice of the speaker (that is, the pitch of the response voice is the pitch of the utterance voice. Dependent on the pitch near the end point). In consideration of the above tendency, the response generation unit 36A of the first embodiment adjusts the pitch of the audio signal Z according to the pitch P of the uttered voice Vx specified by the voice analysis unit 34A, so that the response voice A response signal Y of Vy is generated.

  FIG. 2 is a flowchart of processing executed by the control device 20 of the first embodiment. For example, the process of FIG. 2 is started in response to an instruction from the user U (for example, an instruction to start a voice conversation program) to the voice interaction apparatus 100A.

  When the processing of FIG. 2 is started, the voice acquisition unit 32 waits until the user U starts to pronounce the uttered voice Vx (S10: NO). Specifically, the voice acquisition unit 32 sequentially identifies the volume of the uttered voice Vx by analyzing the utterance signal X supplied from the voice input device 24, and the volume of the uttered voice Vx is set to a predetermined threshold (for example, in advance). It is determined that the utterance voice Vx has started when a state exceeding a fixed value selected in (1) or a variable value according to an instruction from the user U) continues for a predetermined length of time. Note that the detection method of the start of the speech voice Vx (that is, the start point of the speech section) is arbitrary. For example, when the volume of the uttered voice Vx exceeds the threshold and the voice analysis unit 34A detects a significant pitch P, it can be determined that the uttered voice Vx has started.

  When the uttered voice Vx starts (S10: YES), the voice acquisition unit 32 acquires the utterance signal X from the voice input device 24 and stores it in the storage device 22 (S11). The voice analysis unit 34A specifies the pitch P of the uttered voice Vx from the utterance signal X acquired by the voice acquisition unit 32 and stores it in the storage device 22 (S12).

  The voice acquisition unit 32 determines whether or not the user U has finished generating the uttered voice Vx (S13). Specifically, the voice acquisition unit 32 sets the volume of the utterance voice Vx specified from the utterance signal X to a predetermined threshold (for example, a fixed value selected in advance or a variable value according to an instruction from the user U). When the lowering state continues for a predetermined length of time, it is determined that the speech voice Vx has ended. However, a known technique can be arbitrarily employed for detecting the end of the speech voice Vx (that is, the end point of the speech section). As understood from the above description, during the utterance period during which the utterance of the utterance voice Vx is continued (S13: NO), the acquisition of the utterance signal X by the voice acquisition section 32 (S11) and the utterance voice Vx by the voice analysis section 34A. The pitch P is identified (S12).

  As a result of the processing described above, as illustrated in FIGS. 3 and 4, a time series of a plurality of pitches P of the speech voice Vx is specified for the speech section from the start point to the end point tB of the speech voice Vx. In FIG. 3, it is assumed that the user U has pronounced the utterance voice Vx of the question sentence “Is it fun?” That asks the speaker to recognize the emotion, intention, etc. of the utterance partner. In FIG. 4, it is assumed that the user U has pronounced a plain speech utterance Vx that expresses the recognition of the speaker's own emotions, intentions, etc. or requests the utterance partner to agree on the recognition.

  When the uttered voice Vx ends (S13: YES), the response generating unit 36A executes a process (hereinafter referred to as “response generating process”) SA for causing the playback device 26 to reproduce the response voice Vy corresponding to the uttered voice Vx. As described above, the response generation process SA of the first embodiment adjusts the pitch of the voice signal Z in accordance with the pitch P of the uttered voice Vx specified by the voice analysis unit 34A, whereby the response signal of the response voice Vy. This is a process for generating Y.

  FIG. 5 is a flowchart of a specific example of the response generation process SA. As described above, the response generation process SA of FIG. 5 is started when the utterance voice Vx ends (S13: YES). When the response generation process SA is started, the response generation unit 36A, as illustrated in FIG. 3 and FIG. 4, for a section (hereinafter referred to as “tail section”) E including the end point tB of the uttered voice Vx in the uttered voice Vx. The lowest value (hereinafter referred to as “minimum pitch”) Pmin among the plurality of pitches P specified by the voice analysis unit 34A is specified as the prosody of the speech voice Vx (SA1). The tail section E is, for example, a part of the utterance voice Vx extending from the end point tB of the utterance voice Vx to a predetermined length (for example, several seconds) on the near side. As understood from FIG. 3, in the utterance voice Vx of the question sentence, the pitch P tends to increase in the vicinity of the end point tB. Therefore, the pitch P at the minimum point at which the transition of the pitch P of the speech voice Vx changes from a decrease to an increase is specified as the minimum pitch Pmin. On the other hand, as understood from FIG. 4, the pitch P tends to decrease monotonously toward the end point tB in the spoken speech Vx. Therefore, the pitch P at the end point tB of the speech voice Vx is specified as the lowest pitch Pmin.

  The response generation unit 36A generates a response signal Y representing the response voice Vy having a pitch corresponding to the minimum pitch Pmin of the speech voice Vx (SA2). Specifically, as illustrated in FIGS. 3 and 4, the response generation unit 36 </ b> A has the lowest pitch at a specific time point (hereinafter referred to as “target point”) τ on the time axis of the response voice Vy. The response signal Y of the response voice Vy is generated by adjusting the pitch of the voice signal Z so as to coincide with the high Pmin. A preferred example of the target point τ is a start point of a specific mora (typically the last mora) among the plurality of mora constituting the response voice Vy. For example, assuming a voice signal Z of a response voice of “Yes”, as understood from FIGS. 3 and 4, the pitch of the start point of “n” which is the last mora of the voice signal Z is the lowest pitch Pmin. By adjusting (pitch shifting) the pitch over the entire interval of the audio signal Z so as to match the above, the response signal Y of the response audio Vy is generated. In addition, a well-known technique can be arbitrarily employ | adopted for adjustment of a pitch. Further, the target point τ is not limited to the start point of the last mora in the response voice Vy. For example, the pitch can be adjusted with the start point or end point of the response voice Vy as the target point τ.

  When the response signal Y is generated by the above procedure, the response generation unit 36A waits until the time point ty at which playback of the response voice Vy should start (hereinafter referred to as “response playback point”) arrives (SA3: NO). The response reproduction point ty is, for example, a point in time when a predetermined time (for example, 150 ms) has elapsed from the end point tB of the speech voice Vx.

  When the response playback point ty arrives (SA3: YES), the response generator 36A supplies the response signal Y after adjustment according to the minimum pitch Pmin to the playback device 26 to play back the response voice Vy (SA4). . That is, the reproduction of the response voice Vy is started at the response reproduction point ty after a predetermined time has elapsed from the end point tB of the utterance voice Vx. The response generator 36A sequentially supplies the response signal Y from the response playback point ty to the playback device 26 in real time in parallel with the generation (pitch shift) of the response signal Y to play back the response voice Vy. It is also possible. As understood from the above description, the response generation unit 36A of the first embodiment is an element that causes the playback device 26 to play back the response voice Vy having a pitch corresponding to the lowest pitch Pmin in the tail section E of the speech voice Vx. Function.

  When the response generation process SA exemplified above is completed, the control device 20 determines whether or not the user U has instructed the end of the voice dialogue as exemplified in FIG. 2 (S14). If the end of the voice dialogue is not instructed (S14: NO), the process proceeds to step S10. That is, triggered by the start of the uttered voice Vx (S10: YES), the voice acquisition unit 32 acquires the utterance signal X (S11), the voice analysis unit 34A specifies the pitch P (S12), and the response generation unit 36A. The response generation process SA is executed. As understood from the above description, the response voice Vy having a pitch corresponding to the pitch P of the uttered voice Vx is reproduced for each pronunciation of the uttered voice Vx. That is, a voice dialogue is realized in which the pronunciation of an arbitrary utterance voice Vx by the user U and the reproduction of a response voice Vy (for example, a response voice of “yeah”) corresponding to the utterance voice Vx are alternately repeated. . When the end of the voice dialogue is instructed by the user U (S14: YES), the control device 20 ends the process of FIG.

  As described above, in the first embodiment, the response sound Vy having a pitch corresponding to the lowest pitch Pmin in the end section E including the end point tB of the utterance voice Vx is played from the playback device 26. Therefore, it is possible to realize a natural voice conversation simulating the tendency of an actual conversation in which the conversation partner generates a response voice at a pitch corresponding to the pitch near the end point of the uttered voice. Particularly in the first embodiment, since the response voice Vy is reproduced so that the pitch at the start point (target point τ) of the last mora in the response voice Vy matches the minimum pitch Pmin, it is close to an actual dialogue. The effect of realizing a natural voice conversation is particularly remarkable.

<Modification of First Embodiment>
(1) In the first embodiment, the configuration in which the pitch of the target point τ in the response voice Vy is matched with the minimum pitch Pmin in the tail section E of the utterance voice Vx is exemplified. However, the target point τ of the response voice Vy is illustrated. The relationship between the pitch of the voice and the minimum pitch Pmin of the uttered voice Vx is not limited to the above example (a relationship in which both match). For example, the pitch of the response voice Vy at the target point τ can be matched with the pitch obtained by adding or subtracting a predetermined adjustment value (offset) δp to the minimum pitch Pmin. The adjustment value δp is a fixed value selected in advance (for example, a numerical value corresponding to a pitch such as 5 degrees with respect to the lowest pitch Pmin) or a variable value according to an instruction from the user U. Further, according to the configuration in which the adjustment value δp is set to a numerical value corresponding to an integral multiple of the octave, the response voice Vy having a pitch obtained by octave shifting the minimum pitch Pmin is reproduced. It is also possible to switch whether to apply the adjustment value δp according to an instruction from the user U.

(2) In the first embodiment, the pitch of the response voice Vy is controlled according to the pitch P of the utterance voice Vx (specifically, the minimum pitch Pmin of the end section E). The prosody type of the speech voice Vx used for control and the prosody type of the response voice Vy controlled according to the prosody of the speech voice Vx are not limited to the pitch. For example, a configuration that controls the prosody of the response voice Vy according to the volume of the utterance voice Vx (an example of prosody), or a response voice Vy according to the range of the pitch or volume of the utterance voice Vx (an example of the prosody). The structure which controls the prosody of is also adopted. Further, the volume of the response voice Vy (an example of the prosody) is controlled according to the prosody of the uttered voice Vx, and the range of fluctuations in the pitch or volume of the response voice Vy according to the prosody of the uttered voice Vx (other than the prosody) A configuration for controlling the example) may also be employed.

(3) In real human interaction, the prosody of the response speech is not necessarily determined uniformly according to the prosody of the speech speech. That is, the prosody of the response voice tends to depend on the prosody of the utterance voice and can vary for each pronunciation of the utterance voice. Considering the above tendency, the response generation unit 36A can change the prosody (for example, pitch and volume) of the response voice Vy reproduced from the playback device 26 for each utterance voice Vx. Specifically, as described above, in the configuration in which the pitch of the response voice Vy is adjusted to be the pitch obtained by adding or subtracting the adjustment value δp to the minimum pitch Pmin, the response generation unit 36A The adjustment value δp is variably controlled for each sound Vx. For example, the response generation unit 36A generates a random number within a predetermined range for each pronunciation of the uttered voice Vx, and sets the random number as the adjustment value δp. According to the above configuration, it is possible to realize a natural voice conversation simulating the tendency of an actual conversation in which the prosody of the response voice can be changed every time the uttered voice is pronounced.

(4) In the first embodiment, the response signal Y is generated by adjusting the pitch of one type of audio signal Z. However, multiple types of audio signals Z having different pitches are used for generating the response signal Y. It is also possible. For example, a configuration in which the response signal Y is generated by adjusting the pitch of the voice signal Z that is closest to the lowest pitch Pmin of the uttered voice Vx among the plurality of types of voice signals Z can be assumed.

(5) In the first embodiment, the response voice Vy is reproduced from the playback device 26, but the utterance voice Vx is also played back from the playback device 26 by supplying the utterance signal X acquired by the voice acquisition unit 32 to the playback device 26. It is possible. A configuration for switching whether or not to reproduce the uttered voice Vx from the playback device 26 in accordance with an instruction from the user U may be employed.

Second Embodiment
A second embodiment of the present invention will be described. In addition, about the element which an effect | action and function are the same as that of 1st Embodiment in each form illustrated below, the code | symbol used by description of 1st Embodiment is diverted, and each detailed description is abbreviate | omitted suitably.

  FIG. 6 is a configuration diagram of a voice interaction device 100B according to the second embodiment of the present invention. Similarly to the voice interaction apparatus 100A of the first embodiment, the voice interaction apparatus 100B of the second embodiment reproduces a response voice Vy corresponding to the uttered voice Vx produced by the user U. As illustrated in FIG. 6, the voice interaction device 100B of the second embodiment has a configuration in which the response generation unit 36A of the voice interaction device 100A of the first embodiment is replaced with a response generation unit 36B. The configuration and operation of other elements (voice input device 24, playback device 26, voice acquisition unit 32, voice analysis unit 34A) of the voice interactive device 100B are the same as those in the first embodiment.

  In a dialogue between real people, a tendency is observed that the conversation partner pronounces the response speech with a prosody according to the utterance content of the speaker (whether it is a question sentence or a plain sentence). For example, the prosody differs between a response voice for a question sentence and a response voice for a plain text. Specifically, the voice of the answer to the question sentence is relatively loud compared to the voice of the companion sentence, for example, because the answerer's answer (affirmation / denial) needs to be clearly recognized by the speaker. There is a tendency that the volume is pronounced with emphasis on the inflection (time fluctuation of volume or pitch). In consideration of the above tendency, the response generation unit 36B of the second embodiment causes the playback device 26 to play back the prosodic response voice Vy according to the utterance content (question / question distinction) by the utterance voice Vx.

  FIG. 7 illustrates the transition of the pitch P of the utterance voice Vx of the question sentence, and FIG. 8 illustrates the transition of the pitch P of the utterance voice Vx of the plain sentence. As understood from FIGS. 7 and 8, the transition of the pitch P in the vicinity of the end of the utterance voice Vx (temporal in time) between the case where the utterance content of the utterance voice Vx is a question sentence and the case where it is a plain sentence. There is a tendency that the fluctuation tendency) is different. Specifically, the pitch P of the utterance voice Vx of the question sentence is changed from a decrease to an increase or increases monotonously in the end section E as illustrated in FIG. As illustrated in FIG. 8, the high P monotonously decreases from the start point tA to the end point tB of the tail section E. Therefore, by analyzing the transition of the pitch P in the vicinity of the end of the utterance voice Vx (end section E), it is possible to estimate whether the utterance content of the utterance voice Vx corresponds to a question sentence or a plain sentence. It is.

  In consideration of the above tendency, the response generation unit 36B of the second embodiment responds to the prosodic response voice according to the transition of the pitch P in the last section E (ie, the distinction between the question sentence / the plain sentence) in the speech voice Vx. The playback device 26 plays back Vy. Specifically, as illustrated in FIG. 7, when the transition of the pitch P of the utterance voice Vx changes from a decrease to an increase in the end section E, or the pitch P of the utterance voice Vx is monotonous in the end section E. (That is, when the utterance content is estimated to be a question sentence), the prosody response voice Vy suitable for the question sentence is reproduced from the reproduction device 26. On the other hand, as illustrated in FIG. 8, when the pitch P of the speech voice Vx decreases monotonously within the end section E (that is, when the speech content is estimated to be a plain text), it is suitable for a plain text. A prosody response voice Vy is reproduced from the reproduction device 26.

  As illustrated in FIG. 6, the storage device 22 of the voice interaction device 100B of the second embodiment stores a response signal YA and a response signal YB in which the response voice Vy of the specific utterance content is recorded in advance. The response signal YA and the response signal YB have the same utterance content (character notation) but have different prosody. Specifically, the response voice Vy represented by the response signal YA is a voice of “Yes” pronounced with the intention of an affirmative answer to the utterance voice Vx of the question sentence, and the response voice Vy represented by the response signal YB is: This is the voice of “Yun” that is pronounced with the intention of reconciliation with the plain speech utterance voice Vx. Specifically, the response voice Vy of the response signal YA has a prosodic difference that the volume is larger than that of the response voice Vy of the response signal YB, and the range of fluctuations in volume and pitch (ie, inflection) is wide. The response generation unit 36B of the second embodiment selectively supplies one of the response signal YA and the response signal YB stored in the storage device 22 to the playback device 26, so that a plurality of responses with different prosody are obtained. The audio Vy is selectively reproduced. It should be noted that the content of pronunciation can be made different between the response signal YA and the response signal YB.

  FIG. 9 is a flowchart of a response generation process SB for the response generation unit 36B of the second embodiment to cause the playback device 26 to play back the response voice Vy. In the second embodiment, the response generation process SA of FIG. 2 illustrated in the first embodiment is replaced with the response generation process SB of FIG. Processing other than the response generation processing SB is the same as in the first embodiment. The response generation process SB in FIG. 9 is started when the utterance voice Vx ends (S13: YES).

  When the response generation process SB is started, the response generation unit 36B has an average (hereinafter referred to as “first average pitch”) Pave1 of a plurality of pitches P in the first interval E1 in the end interval E of the speech voice Vx, An average of a plurality of pitches P in the second section E2 (hereinafter referred to as “second average pitch”) Pave2 is calculated (SB1). As illustrated in FIGS. 7 and 8, the first section E1 is a front section (for example, a section including the start point tA of the end section E) of the end section E, and the second section E2 is the end section E. Of these, it is a section behind the first section E1 (for example, a section including the end point tB of the end section E). Specifically, the first half of the tail section E is defined as the first section E1, and the second half of the tail section E is defined as the second section E2. However, the conditions of the first section E1 and the second section E2 are not limited to the above examples. For example, a configuration in which the first interval E1 and the second interval E2 are moved back and forth with an interval, or a configuration in which the time lengths are different between the first interval E1 and the second interval E2 can be adopted.

  The response generator 36B compares the first average pitch Pave1 in the first section E1 with the second average pitch Pave2 in the second section E2, and determines whether the first average pitch Pave1 is lower than the second average pitch Pave2. It is determined whether or not (SB2). As described above, the transition of the pitch P of the utterance voice Vx of the question sentence tends to change from a decrease to an increase or monotonously increase in the end section E. Therefore, as illustrated in FIG. 7, the first average pitch Pave1 is likely to be lower than the second average pitch Pave2 (Pave1 <Pave2). On the other hand, the pitch P of the plain speech utterance voice Vx tends to decrease monotonously in the end section E. Therefore, as illustrated in FIG. 8, the first average pitch Pave1 is likely to exceed the second average pitch Pave2 (Pave1> Pave2).

  In consideration of the above tendency, when the first average pitch Pave1 is lower than the second average pitch Pave2 (SB2: YES), that is, when the possibility that the speech Vx is a questionable sentence is high, the second The response generation unit 36B of the embodiment selects the response signal YA corresponding to the response voice Vy of the answer to the question sentence from the storage device 22 (SB3). On the other hand, when the first average pitch Pave1 exceeds the second average pitch Pave2 (SB2: NO), that is, when there is a high possibility that the speech voice Vx is a plain text, the response generation unit 36B Is selected from the storage device 22 (SB4).

  When the response signal Y (Y1, Y2) corresponding to the transition of the pitch P of the uttered voice Vx is selected by the above procedure, the response generator 36B arrives at the response reproduction point ty (SB5) as in the first embodiment. : YES), the response signal Vy is reproduced by supplying the response signal Y to the reproduction device 26 (SB6). Specifically, when the pitch P of the utterance voice Vx changes from a decrease to an increase within the end section E, or when the pitch P of the utterance voice Vx increases monotonically within the end section E (SB2: YES). When the answer voice Vy of the answer to the question sentence is reproduced and the pitch P of the utterance voice Vx decreases monotonously in the end section E (SB2: NO), the answer voice Vy of the consent to the plain text is reproduced. . That is, the prosody of the response voice Vy reproduced from the playback device 26 differs between the case where the utterance voice Vx is a question sentence and the case where it is a plain sentence.

  The acquisition of the speech signal X by the voice acquisition unit 32 (S11), the specification of the pitch P by the voice analysis unit 34A (S12), and the response generation process SB by the response generation unit 36B are as follows. (S14: NO). Therefore, as in the first embodiment, a voice conversation is realized in which the sound of an arbitrary uttered voice Vx by the user U and the reproduction of the response voice Vy for the uttered voice Vx are alternately repeated.

  As described above, in the second embodiment, the prosody response voice Vy corresponding to the transition of the pitch P in the last section E of the speech voice Vx is played from the playback device 26. Therefore, it is possible to realize a natural voice conversation simulating the tendency of an actual conversation in which the conversation partner sounds a response voice with a prosody according to the utterance content of the speaker. In the second embodiment, in particular, when the transition of the pitch P changes from a decrease to an increase within the end section E, or when the pitch P increases monotonically within the end section E, and the end point from the start point tA of the end section E. Since the prosody of the response voice Vy is different from the case where the pitch P decreases monotonously over tB, a natural voice dialog simulating the actual dialog tendency that the prosody of the response voice is different between the question sentence and the plain text. Can be realized.

  Further, in the second embodiment, the prosody of the response voice Vy according to the result of comparing the first average pitch Pave1 in the first section E1 and the second average pitch Pave2 in the second section E2 in the end section E. Therefore, there is an advantage that transition of the pitch P can be evaluated by simple processing of averaging and comparing a plurality of pitches P (and thus the prosody of the response voice Vy can be selected).

<Modification of Second Embodiment>
(1) In the second embodiment, a plurality of response signals Y (YA, YB) stored in advance in the storage device 22 are selectively supplied to the playback device 26. However, a single response signal recorded in advance is used. By adjusting Y, the response generation unit 36B can generate a prosodic response signal Y corresponding to the transition of the pitch P in the last section E of the speech voice Vx. For example, assuming a configuration in which the response signal YA of the response voice Vy to the plain text is held in the storage device 22, the response generation unit 36B increases the volume of the response signal YA and the volume when the uttered voice Vx is a question sentence. In addition, the response signal YB of the response voice Vy of the answer is generated by expanding the range of fluctuation of the pitch, while the response signal YA is supplied to the playback device 26 when the utterance voice Vx is a plain text. Note that it is also possible to generate the response signal YA of the response voice Vy of the consent to the plain text by reducing the volume of the initial response signal Y and reducing the range of fluctuation of the volume and pitch.

  According to the configuration in which the response signal Y having different prosody is generated by adjusting the one response signal Y, it is not necessary to store a plurality of response signals Y (YA, YB) having different prosody in the storage device 22. There is an advantage that the storage capacity required for the storage device 22 is reduced. On the other hand, according to the configuration of the second embodiment that selectively uses a plurality of response signals Y having different prosody, there is no need to adjust the initial prosody of the response signal Y according to the utterance content of the utterance voice Vx. Therefore, there is an advantage that the processing load of the response generation unit 36B is reduced.

(2) In the second embodiment, the first average pitch Pave1 in the first section E1 and the second average pitch Pave2 in the second section E2 in the end section E are compared, but the utterance of the speech voice Vx is compared. The method for estimating whether the content corresponds to a question sentence or a plain text is not limited to the above examples. For example, in the plain speech utterance voice Vx, since the pitch P decreases monotonously in the end section E, the pitch P tends to become the minimum pitch Pmin at the end point tB of the end section E. Therefore, when the time length of the rear section of the end section E at which the pitch P becomes the minimum pitch Pmin is sufficiently shorter than the front section (for example, when it is below a predetermined threshold), the speech voice It is also possible to estimate that the utterance content of Vx corresponds to a plain text. It is also possible to estimate whether the utterance content of the utterance voice Vx corresponds to the question sentence or the plain sentence according to the transition of the pitch P before and after the point of the lowest pitch Pmin in the end section E. is there. For example, when the pitch P rises after the time of the lowest pitch Pmin in the end section E, the response generation unit 36B estimates that the utterance content of the uttered voice Vx corresponds to the question sentence.

<Third Embodiment>
FIG. 10 is a configuration diagram of a voice interactive apparatus 100C according to the third embodiment of the present invention. Similar to the voice interaction apparatus 100A of the first embodiment, the voice interaction apparatus 100C of the third embodiment reproduces a response voice Vy corresponding to the uttered voice Vx pronounced by the user U. In the third embodiment, in addition to an answer to the utterance voice Vx or a response voice (hereinafter referred to as “second response voice”) Vy2, a response voice (hereinafter referred to as “first response voice”) indicating a response to the utterance voice Vx. ) Vy1 can be regenerated from the regenerator 26. The first response voice Vy1 is a voice such as “E?” Or “What?” For rehearsing the utterance voice Vx to the speaker. As illustrated in FIG. 10, the storage device 22 of the voice interaction device 100C of the third embodiment includes a response signal Y1 containing the first response voice Vy1 for answering questions and a phase other than answering questions (for example, “Yes”). The response signal Y2 recorded with the second response voice Vy2 of (ii) is stored.

  As illustrated in FIG. 10, the voice interaction device 100C of the third embodiment replaces the voice analysis unit 34A and the response generation unit 36A of the voice interaction device 100A of the first embodiment with the voice analysis unit 34C and the response generation unit 36C. This is a replacement configuration. The configuration and operation of the other elements (voice input device 24, playback device 26, voice acquisition unit 32) of the voice interactive device 100C are the same as in the first embodiment.

  The voice analysis unit 34C of the third embodiment specifies the prosody index value Q from the utterance signal X acquired by the voice acquisition unit 32. The prosodic index value Q is an index value related to the prosody of the utterance voice Vx, and is calculated for each utterance voice Vx (for each unit when a series of utterances from the start point to the end point of the utterance voice Vx is used as a unit). Specifically, the average value of pitches, the fluctuation range of pitches, the average value of volume, or the fluctuation range of volume in the utterance section of the utterance voice Vx is calculated from the utterance signal X as the prosodic index value Q. As described above, the response generation unit 36C of the third embodiment causes the playback device 26 to selectively play back the first response voice Vy1 that represents the answer to the uttered voice Vx and the second response voice Vy2 other than the answer.

  In a dialogue between real people, when the prosody of the uttered voice of the speaker changes, there is a tendency that it becomes difficult for the conversation partner to listen to the uttered voice and the necessity of answering questions increases. Specifically, when the prosody of the utterance of the speaker deviates from the tendency of the utterance of the speaker in the past (for example, the volume of the actual utterance is lower than the volume assumed by the conversation partner from the past tendency) ), The conversation partner cannot properly listen to the uttered voice, and as a result, there is a high possibility that a question is returned to the speaker. Considering the above tendency, the response generation unit 36C of the third embodiment compares the prosodic index value Q specified by the speech analysis unit 34C with the threshold value QTH, and determines the first response speech Vy1 and the first response speech Vy1 according to the comparison result. Either of the two response voices Vy2 is caused to be played back by the playback device 26. The threshold value QTH is set to a representative value (for example, an average value) of the prosodic index value Q of the speech voice Vx spoken by the user U in the past. That is, the threshold value QTH corresponds to a standard prosody estimated from the user U's past speech. When the prosodic index value Q of the utterance voice Vx deviates from the threshold value QTH, the first response voice Vy1 is returned. When the prosodic index value Q approximates the threshold value QTH, the second response is a conflict. Audio Vy2 is reproduced.

  FIG. 11 is a flowchart of processing executed by the control device 20 of the third embodiment. For example, the process in FIG. 11 is started in response to an instruction from the user U (for example, an instruction to start a program for voice conversation) to the voice conversation apparatus 100C.

  As in the first embodiment, when the utterance voice Vx is started (S20: YES), the voice acquisition unit 32 acquires the utterance signal X from the voice input device 24 and stores it in the storage device 22 (S21). The voice analysis unit 34C specifies the feature quantity q related to the prosody of the uttered voice Vx from the utterance signal X acquired by the voice acquisition unit 32 (S22). The feature quantity q is, for example, the pitch P or volume of the speech voice Vx. The acquisition of the utterance signal X by the voice acquisition unit 32 (S21) and the specification of the feature quantity q by the voice analysis unit 34C (S22) are repeated until the end of the utterance voice Vx (S23: NO). That is, a time series of a plurality of feature quantities q of the utterance voice Vx is specified for the utterance section from the start point to the end point tB of the utterance voice Vx.

  When the uttered voice Vx ends (S23: YES), the voice analysis unit 34C calculates the prosodic index value Q from the time series of a plurality of feature quantities q specified for the utterance section from the start point to the end point of the uttered voice Vx (S24). ). Specifically, the speech analysis unit 34C calculates an average value or a variation range (range) of the plurality of feature quantities q in the utterance section as the prosodic index value Q.

  When the prosody index value Q of the current speech voice Vx is calculated by the process described above, the response generation unit 36C executes the response generation process SC for causing the playback device 26 to play back the response voice Vy. The response generation process SC of the third embodiment is a process in which the playback device 26 selectively plays back either the first response voice Vy1 or the second response voice Vy2 according to the prosodic index value Q calculated by the voice analysis unit 34C. It is.

When the response generation process SC is completed, the voice analysis unit 34C updates the threshold value QTH according to the prosodic index value Q of the uttered voice Vx (S25). Specifically, the voice analysis unit 34C calculates a representative value (for example, an average value or a median value) of a plurality of prosodic index values Q of the past utterance voice Vx including the utterance voice Vx as the updated threshold value QTH. . For example, as expressed by the following formula (1), the weighted average (exponential moving average) between the current prosodic index value Q and the threshold value QTH before update is calculated as the threshold value QTH after update. The symbol α in the formula (1) is a predetermined positive number (forgetting factor) less than 1.
QTH = α ・ Q + (1-α) QTH …… (1)
As understood from the above description, the speech analysis unit 34C of the third embodiment functions as an element that sets the representative value of the prosodic index value Q in a plurality of past speech speech Vx as the threshold value QTH. The threshold value QTH is updated to a numerical value reflecting the prosodic index value Q of the utterance voice Vx for each pronunciation of the utterance voice Vx, and becomes a numerical value corresponding to a standard prosody estimated from the utterance of the user U over a plurality of times. . However, it is also possible to fix the threshold value QTH to a predetermined value. For example, the average value of the prosodic index values Q specified from the uttered voices of an unspecified number of speakers can be set as the threshold value QTH.

  Acquisition of speech signal X by voice acquisition unit 32 (S21), calculation of prosodic index value Q by voice analysis unit 34C (S22, S24), response generation processing SC by response generation unit 36C, and threshold by voice analysis unit 34C The update of QTH (S25) is repeated for each pronunciation of the uttered voice Vx until the end of the voice conversation is instructed by the user U (S26: NO). Accordingly, a voice dialogue is realized in which the pronunciation of the uttered voice Vx by the user U and the selective reproduction of the first response voice Vy1 (question answer) and the second response voice Vy2 (interaction) are alternately repeated. .

  FIG. 12 is a flowchart of the response generation process SC of the third embodiment. When the response generation process SC is started, the response generation unit 36C compares the prosodic index value Q specified by the speech analysis unit 34C with the current threshold value QTH, and a predetermined range including the threshold value QTH (hereinafter referred to as “allowable range”). It is determined whether or not the prosodic index value Q is included in R (SC1). 13 and 14 illustrate the transition of the feature quantity q specified by the voice analysis unit 34C from the speech voice Vx. As illustrated in FIGS. 13 and 14, the allowable range R is a range having a predetermined width with the threshold value QTH being a median value. The process of comparing the prosodic index value Q and the threshold value QTH (SC1) determines whether or not the absolute value of the difference between the prosodic index value Q and the threshold value QTH exceeds a predetermined value (for example, half the range width of the allowable range R). It can also be realized as a determination process.

  In FIG. 13, it is assumed that the prosodic index value Q is a numerical value inside the allowable range R. The inclusion of the prosodic index value Q within the allowable range R means that the prosody of the utterance voice Vx of this time approximates the standard prosody of the user U (past utterance tendency). In other words, assuming a dialogue between real people, it can be evaluated that the conversation partner is easy to listen to the uttered voice (a situation where there is a low possibility that a question is returned to the speaker). Therefore, when the prosodic index value Q is a numerical value within the allowable range R (SC1: YES), the response generation unit 36C sends the response signal Y2 of the second response voice Vy2 that is in conflict with the utterance voice Vx from the storage device 22. Select (SC2).

  On the other hand, in FIG. 14, it is assumed that the prosodic index value Q is a numerical value outside the allowable range R (specifically, a numerical value lower than the lower limit value of the allowable range R). The fact that the prosodic index value Q is not included in the allowable range R means that the prosody of the uttered voice Vx is deviated from the standard prosody of the user U. In other words, assuming a dialogue between actual humans, it can be evaluated that the conversation partner is difficult to hear the uttered voice (a situation where there is a high possibility that it is necessary to answer the speaker). Therefore, when the prosodic index value Q is a numerical value outside the allowable range R (SC1: NO), the response generation unit 36C causes the second response voice Vy1 (for example, “??” “what” to answer the uttered voice Vx). Is selected from the storage device 22 as an object to be supplied to the playback device 26 (SC3).

  When the response signal Y (response target response voice Vy) corresponding to the prosodic index value Q is selected by the above procedure, the response generator 36C arrives at the response playback point ty (SC4: YES) as in the first embodiment. ), The response signal Vy (the first response sound Vy1 or the second response sound Vy2) is reproduced by supplying the response signal Y to the reproduction device 26 (SC5). That is, when the prosodic index value Q is included in the allowable range R, the second response voice Vy2 is reproduced, and when the prosodic index value Q is not included in the allowable range R, the answering first response voice Vy2 is reproduced. Vy1 is played back.

  As described above, in the third embodiment, the first response voice Vy1 that indicates a response to the uttered voice Vx and the second response voice Vy2 other than the response are selectively played back from the playback device 26. Therefore, it is possible to realize a natural voice dialogue that simulates the tendency of an actual dialogue in which not only the talker's utterance but also the questioning / replying (rehearsal) to the talker occurs appropriately.

  In the third embodiment, since either the first response voice Vy1 or the second response voice Vy2 is selected according to the result of comparing the prosody index value Q representing the prosody of the utterance voice Vx with the threshold value QTH, the utterance It is possible to realize a natural voice conversation that simulates the tendency of an actual conversation that when the prosody of the voice changes unexpectedly, it becomes difficult to listen and the necessity of answering questions increases. Particularly in the third embodiment, since the representative value of the prosodic index value Q over a plurality of past utterances Vx is set as the threshold value QTH, the prosody of the utterance of the utterer is the standard prosody of the utterer (that is, the dialog) There is also an advantage that a natural voice conversation simulating the tendency of an actual conversation that the question is likely to be returned from the conversation partner when it deviates from the prosody of the partner). Moreover, the first response voice Vy1 is selected when the prosodic index value Q is a numerical value outside the allowable range R including the threshold value QTH, and the second response voice Vy2 is selected when the numerical value is inside the allowable range R. Therefore, for example, the first response voice Vy1 is excessively frequently compared with the configuration in which the first response voice Vy1 and the second response voice Vy2 are selected only in accordance with the magnitude relationship between the prosodic index value Q and the threshold value QTH. Can be reduced (the first response voice Vy1 is reproduced at an appropriate frequency).

<Modification of Third Embodiment>
In the third embodiment, the reproduction of the first response voice Vy1 and the reproduction of the second response voice Vy2 are selected according to the prosodic index value Q of the utterance voice Vx. However, a predetermined frequency is used regardless of the characteristics of the utterance voice Vx. It is also possible to play back the first response voice Vy1 asking. Specifically, the response generation unit 36C plays back to the playback device 26 the first response voice Vy1 that answers the utterance voice Vx that is randomly selected from the plurality of utterance voices Vx that the user U sequentially generates. On the other hand, for the remaining utterance voice Vx, the corresponding second response voice Vy2 is reproduced. For example, the response generation unit 36C generates a random number within a predetermined range for each utterance of the utterance voice Vx, and selects the first response voice Vy1 when the random number exceeds the threshold, while the random number falls below the threshold. In this case, the second response voice Vy2 is selected. In the modification illustrated above, since the first response voice Vy1 for answering the utterance voice Vx randomly selected from the plurality of utterance voices Vx is reproduced, the answer to the utterance voice is randomly generated. It is possible to realize a natural voice conversation simulating the tendency of actual voice conversation.

  In the above configuration, the response generation unit 36C can variably set the ratio of the number of times the first response voice Vy1 is reproduced to the number of times the utterance voice Vx is uttered (that is, the reproduction frequency of the first response voice Vy1). For example, the response generation unit 36C controls the reproduction frequency of the first response voice Vy1 by adjusting the threshold value to be compared with the random number. For example, when the reproduction frequency of the first response voice Vy1 is set to 30%, the first response voice Vy1 is reproduced for 30% of the total number of utterances of the utterance voice Vx, and the remaining number of times is 70%. The second response voice Vy2 is reproduced for the utterance. The reproduction frequency of the first response voice Vy1 (for example, a threshold value compared with a random number) is variably set according to an instruction from the user U, for example.

<Fourth embodiment>
FIG. 15 is a configuration diagram of a voice interaction device 100D according to the fourth embodiment of the present invention. Similar to the voice interaction apparatus 100A of the first embodiment, the voice interaction apparatus 100D of the fourth embodiment reproduces a response voice Vy corresponding to the uttered voice Vx pronounced by the user U.

  As illustrated in FIG. 15, the voice interaction device 100D of the fourth embodiment replaces the voice analysis unit 34A and the response generation unit 36A of the voice interaction device 100A of the first embodiment with the history management unit 38 and the response generation unit 36D. This is a replacement configuration. The configuration and operation of other elements (voice input device 24, playback device 26, and voice acquisition unit 32) of the voice interactive device 100D are the same as those in the first embodiment. The storage device 22 of the fourth embodiment stores a response signal Y representing the response voice Vy of specific utterance content. In the following description, a response voice Vy of “Yes” that means a conflict with the speech voice Vx is illustrated.

  The history management unit 38 in FIG. 15 generates a history (hereinafter referred to as “use history”) H of a voice conversation by the voice interaction device 100D. The usage history H of the fourth embodiment is the number of voice conversations executed in the past (hereinafter referred to as “use count”) N using the voice interaction device 100D. Specifically, the voice dialogue is started once (starting of the voice dialogue device 100D) until ending (that is, one voice dialogue including a plurality of pairs of the utterance voice Vx and the response voice Vy). As a result, the history management unit 38 counts the number of voice conversations as the number of uses N. The usage history H generated by the history management unit 38 is stored in the storage device 22.

  The response generation unit 36D of the fourth embodiment causes the playback device 26 to play back the prosodic response voice Vy corresponding to the usage history H generated by the history management unit 38. That is, the prosody of the response voice Vy is variably controlled according to the usage history H. In the fourth embodiment, the reproduction standby time W of the response voice Vy is controlled according to the usage history H as the prosody of the response voice Vy. The waiting time W is the length of time from the end point tB of the utterance voice Vx to the response playback point ty of the response voice Vy (that is, the interval between the utterance voice Vx and the response voice Vy).

  In actual human-to-human dialogue, a tendency is observed in which the prosody of the uttered speech changes with time as the dialogue with a specific dialogue partner is repeated. Specifically, at the stage immediately after starting the first face-to-face conversation (the stage where each person is not used to the conversation with the conversation partner), both parties cannot grasp a suitable interval etc. specific to the conversation partner. There is a tendency that the time from the utterance to the response to the utterance is long (that is, the conversation is awkward), and the time is shortened (that is, the conversation can be performed with good tempo) as the conversation with the conversation partner is repeated. Considering the above tendency, the response generation unit 36D of the fourth embodiment has a longer waiting time W for the response voice Vy when the usage history N indicated by the usage history H is greater than when the usage count N is small. The waiting time W is controlled in accordance with the usage history H so as to be shortened.

  FIG. 16 is a flowchart of processing executed by the control device 20 of the fourth embodiment. For example, the process of FIG. 16 is started in response to an instruction from the user U (instruction to start a voice conversation program) to the voice interaction apparatus 100D. At the stage when the voice dialogue by the voice dialogue device 100D is first started, the usage history H is set to an initial value (for example, N = 0).

  As in the first embodiment, when the utterance voice Vx is started (S30: YES), the voice acquisition unit 32 acquires the utterance signal X from the voice input device 24 and stores it in the storage device 22 (S31). The acquisition of the utterance signal X by the voice acquisition unit 32 is repeated until the end of the utterance voice Vx (S32: NO).

  When the speech voice Vx ends (S32: YES), the response generation unit 36D performs a response generation process SD for causing the playback device 26 to play back the response voice Vy of the prosody according to the usage history H stored in the storage device 22. Run. As described above, the response generation process SD of the fourth embodiment is a process for controlling the waiting time W from the end point tB of the utterance voice Vx to the response playback point ty where the playback of the response voice Vy is started according to the usage history H. It is. The acquisition of the speech signal X by the voice acquisition unit 32 (S31) and the response generation process SD by the response generation unit 36D are repeated until the end of the voice conversation is instructed by the user U (S33: NO). Therefore, as in the first embodiment, a voice conversation is realized in which the sound of an arbitrary uttered voice Vx by the user U and the reproduction of the response voice Vy for the uttered voice Vx are alternately repeated.

  When the end of the voice dialogue is instructed by the user U (S33: YES), the history management unit 38 updates the usage history H stored in the storage device 22 to the content that takes into account the current voice dialogue (S34). ). Specifically, the history management unit 38 increases the usage count N indicated by the usage history H by one. Accordingly, the usage history H increases by 1 each time a voice conversation is performed by the voice conversation device 100D. After the usage history H is updated, the process of FIG.

  FIG. 17 is a flowchart of the response generation process SD of the fourth embodiment, and FIGS. 18 and 19 are explanatory diagrams of the response generation process SD. When the response generation process SD is started, the response generation unit 36D variably sets the standby time W according to the usage history H stored in the storage device 22 (SD1 to SD3). Specifically, the response generation unit 36D first determines whether or not the usage count N indicated by the usage history H exceeds a predetermined threshold value NTH (SD1). When the number of uses N exceeds the threshold value NTH (SD1: YES), the response generation unit 36D sets a predetermined basic value w0 (for example, 150 ms) as the standby time W as illustrated in FIG. 18 (SD2). On the other hand, when the number of uses N is less than the threshold value NTH (SD1: NO), the response generator 36D, as exemplified in FIG. 19, adds the predetermined adjustment value (offset) δw to the basic value w0 (w0 + δw). Is set as the waiting time W (SD3). The adjustment value δw is set to a predetermined positive number. In the above description, the standby time W is binary controlled in accordance with whether or not the usage count N exceeds the threshold value NTH. However, the standby time W is multivalued or continuously in accordance with the usage count N. It is also possible to change.

  The response generation unit 36D waits until the standby time W set according to the usage history H in the above processing has elapsed from the end point tB of the uttered voice Vx (SD4: NO). When the response playback point ty comes due to the elapse of the standby time W (SD4: YES), the response generation unit 36D supplies the response signal Y stored in the storage device 22 to the playback device 26 to play back the response voice Vy. (SD5). As understood from the above description, the response generation unit 36D of the fourth embodiment reproduces the response voice Vy of the prosody (the waiting time W in the fourth embodiment) according to the usage history H of the voice interaction apparatus 100D. To play. Specifically, when the usage history N indicated by the usage history H is large, the response voice Vy is reproduced as the standby time W of the basic value w0 elapses, and when the usage count N is small, the response voice Vy is adjusted to the basic value w0. The response voice Vy is reproduced as the standby time W obtained by adding the value δw elapses. That is, the standby time W is shortened when the number of uses N is large.

  As described above, in the fourth embodiment, since the response voice Vy of the prosody (waiting time W) corresponding to the voice conversation use history H by the voice conversation device 100D is reproduced, the conversation with the specific partner is repeated. It is possible to realize a natural voice conversation that simulates the tendency of an actual conversation in which the prosody of the spoken voice changes with time. In the fourth embodiment, in particular, the standby time W that is the interval between the utterance voice Vx and the response voice Vy is controlled according to the usage history H. Therefore, at the stage immediately after the start of the dialogue at the first meeting, the interval between the utterance and the response is long, and the natural conversation that simulates the tendency of the actual dialogue that the interval is shortened as the dialogue with the dialogue partner is repeated. Spoken dialogue is realized.

<Modification>
The voice interaction apparatus 100 (100A, 100B, 100C, 100D) exemplified in the above-described embodiments can be variously modified. Specific modifications are exemplified below. Two or more modes arbitrarily selected from the following examples can be appropriately combined within a range that does not contradict each other.

(1) It is possible to combine any two or more configurations selected from the first to fourth embodiments. Specifically, the configuration of the first embodiment that controls the prosody of the response voice Vy according to the prosody of the speech voice Vx (for example, pitch P) is similarly applied to the second to fourth embodiments. obtain. For example, in the second embodiment, the prosody of the response signal Y selected in step SB3 or step SB4 in FIG. 9 is controlled according to the prosody of the uttered voice Vx (for example, pitch P) and reproduced from the playback device 26. Is also possible. Similarly, in the third embodiment, a configuration is adopted in which the prosody of the response signal Y selected in step SC2 or step SC3 in FIG. 12 is controlled according to the prosody of the speech voice Vx. In the fourth embodiment, in FIG. A configuration may be adopted in which the prosody of the response signal Y acquired from the storage device 22 in step SD5 is controlled according to the prosody of the speech voice Vx. In the configuration in which the first embodiment is applied to the second to fourth embodiments, as in the first embodiment, for example, at the start point of a specific mora (typically the last mora) of the response voice Vy. The pitch of the response signal Y is adjusted so that the pitch matches the lowest pitch Pmin in the last section E of the speech voice Vx.

  The configuration of the third embodiment that selectively reproduces the first response voice Vy1 that answers the uttered voice Vx and the second response voice Vy2 that does not answer the question may also be applied to each form other than the third embodiment. Is possible. Further, the configuration of the fourth embodiment for controlling the prosody of the response voice Vy (for example, the waiting time W) according to the voice conversation use history H can be applied to the first to third embodiments. .

(2) Various variables related to the above-described forms of voice conversation are variably set according to instructions from the user U, for example. For example, the playback volume of the response voice Vy is controlled according to an instruction from the user U, and the actual playback is performed among a plurality of types of response voices Vy having different gender or voice quality (friendly voice, severe voice) of the speaker. A configuration in which the type of response voice Vy reproduced from the device 26 is selected according to an instruction from the user U can also be adopted. In the first to third embodiments, the length of the waiting time W from the end point tB of the utterance voice Vx to the response playback point ty of the response voice Vy may be set according to an instruction from the user U. Is possible.

(3) In the modified example of the third embodiment, the reproduction frequency of the first response voice Vy1 for answering the uttered voice Vx is variably set according to the instruction from the user U, but other than the instruction from the user U It is also possible to control the reproduction frequency of the first response voice Vy1 according to the above factors. Specifically, a configuration may be employed in which the response generation unit 36D of the third embodiment controls the reproduction frequency of the first response voice Vy1 according to the usage history H exemplified in the fourth embodiment. For example, in a conversation between real people, the characteristics of the utterance (such as moustache and tone) of the conversation partner can be grasped as the conversation with a specific conversation partner is repeated, and as a result, the frequency of answering the spoken speech decreases. The tendency to do is assumed. Considering the above tendency, a configuration in which the reproduction frequency of the first response voice Vy1 is reduced as the number of uses N indicated by the use history H increases.

(4) In the fourth embodiment, the usage count N of the voice conversation is exemplified as the usage history H. However, the usage history H is not limited to the usage count N. For example, the number of times the response voice Vy is reproduced once in the voice conversation, the frequency of use of the voice conversation (the number of uses per unit time), the period of use of the voice conversation (for example, from the first use of the voice conversation apparatus 100) Elapsed time), and the elapsed time since the last use of the voice interaction apparatus 100 can be applied to the control of the standby time W as the usage history H.

(5) In the first embodiment, the response signal Y is generated and reproduced from the audio signal Z stored in advance in the storage device 22. In the second to fourth embodiments, the response signal Y is stored in advance in the storage device 22. The response signal Y representing the response voice Vy of the specific utterance content can be synthesized by, for example, a known voice synthesis technique. For synthesizing the response signal Y, for example, speech synthesis using a unit connection type speech synthesis or a statistical model such as a hidden Markov model is preferably used. Further, the speech voice Vx and the response voice Vy are not limited to human voices. For example, it is possible to use an animal cry as the utterance voice Vx or the response voice Vy.

(6) In each of the above-described embodiments, the configuration in which the voice interaction device 100 includes the voice input device 24 and the playback device 26 has been exemplified. However, the voice dialogue device 100 can be connected to a device (voice input / output device) separate from the voice interaction device 100. It is also possible to install an input device 24 and a playback device 26. The voice interaction device 100 is realized by a terminal device such as a mobile phone or a smartphone, and the voice input / output device is realized by an electronic device such as an animal-type toy or a robot. The voice interactive device 100 and the voice input / output device can communicate with each other wirelessly or by wire. That is, the speech signal X generated by the voice input device 24 of the voice input / output device is transmitted to the voice dialog device 100 wirelessly or by wire, and the response signal Y generated by the voice dialog device 100 is transmitted from the voice input device of the voice input device wirelessly or by wire. It is transmitted to the playback device 26.

(7) In each of the above-described embodiments, the voice interactive device 100 is realized by an information processing device such as a mobile phone or a personal computer. However, a part or all of the functions of the voice interactive device 100 are performed by a server device (so-called cloud server). It can also be realized. Specifically, the voice interactive device 100 is realized by a server device that communicates with a terminal device via a communication network such as a mobile communication network or the Internet. For example, the voice interaction apparatus 100 receives the utterance signal X generated by the voice input device 24 of the terminal apparatus from the terminal apparatus, and generates the response signal Y from the utterance signal X with the configuration according to each of the above-described embodiments. Then, the voice interaction device 100 transmits a response signal Y generated from the speech signal X to the terminal device, and causes the playback device 26 of the terminal device to play back the response voice Vy. The voice interactive device 100 is realized by a single device or a set of a plurality of devices (that is, a server system). In addition, some functions of the voice interaction device 100 according to the above-described embodiments (for example, the voice acquisition unit 32, the voice analysis units 34A and 34C, the response generation units 36A, 36B, 36C, and 36D, and the history management unit 38) ) Can be realized by the server device, and other functions can be realized by the terminal device. Whether each function realized by the voice interactive apparatus 100 is realized by the server apparatus or the terminal apparatus (function sharing) is arbitrary.

(8) In each of the above-described embodiments, the response voice Vy having a specific utterance content (for example, “Yes”) is reproduced with respect to the utterance voice Vx, but the utterance content of the response voice Vy is limited to the above examples. Not. For example, the utterance content of the utterance voice Vx is analyzed by voice recognition and morphological analysis for the utterance signal X, and a response voice Vy having an appropriate content for the utterance content is selected from a plurality of candidates and reproduced by the reproduction device 26. Is also possible. Note that in a configuration in which voice recognition and morphological analysis are not performed (for example, examples in the first to fourth embodiments), the response voice Vy of the utterance content prepared in advance is reproduced regardless of the utterance voice Vx. Therefore, when considered simply, it can be inferred that a natural dialogue is not established, but in reality, humans can interact with each other by controlling the prosody of the response voice Vy in various ways as illustrated in the above examples. The user U can perceive a feeling like a natural dialogue. On the other hand, according to the configuration in which voice recognition and morphological analysis are not performed, there is an advantage that processing delay and processing load due to these processes are reduced or eliminated.

(9) The spoken dialogue apparatus 100 (100A, 100B, 100C, 100D) exemplified in the above-described embodiments can be used for evaluating dialogue between actual humans. For example, when the prosody of response speech (hereinafter referred to as “observation speech”) observed in an actual human dialogue is compared with the prosody of response speech Vy generated in the above-described form, the prosody is similar between the two While the observation speech is evaluated as appropriate, the observation speech can be evaluated as inappropriate when the prosody diverges between the two. The apparatus (dialog evaluation apparatus) that performs the evaluation exemplified above can be used for training of dialogue between humans.

(10) The voice interaction device 100 (100A, 100B, 100C, 100D) exemplified in the above-described embodiments can be realized by the cooperation of the control device 20 and the program for voice interaction as described above. The program for voice interaction can be provided in a form stored in a computer-readable recording medium and installed in the computer. The recording medium is, for example, a non-transitory recording medium, and an optical recording medium (optical disk) such as a CD-ROM is a good example, but a known arbitrary one such as a semiconductor recording medium or a magnetic recording medium This type of recording medium can be included. It is also possible to distribute the program to a computer in the form of distribution via a communication network. The present invention can also be realized as an operation method (voice interaction method) of the voice interaction apparatus 100 exemplified in the above-described embodiments. A computer (speech dialogue apparatus 100) that is an operation subject of the voice dialogue method is, for example, a single computer or a system constituted by a plurality of computers.

100 (100A, 100B, 100C, 100D) …… Voice interaction device, 20 …… Control device, 22 …… Storage device, 24 …… Voice input device, 242 …… Sound collecting device, 244 …… A / D converter , 26... Playback device, 262... D / A converter, 264 .. sound emitting device, 32... Voice acquisition unit, 34 A, 34 C .. speech analysis unit, 36 A, 36 B, 36 C, 36 D. 38, history management unit.

Claims (8)

An audio acquisition unit for acquiring an utterance signal representing the utterance voice;
A voice interaction device comprising: a response generation unit that causes a playback device to selectively play back a first response voice that represents a response to the spoken voice and a second response speech other than the response to the question. Comprising a speech analysis unit for identifying a prosodic index value representing the prosody of the speech speech from the speech signal;
The spoken dialogue according to claim 1, wherein the response generation unit compares a prosodic index value of the uttered voice with a threshold value, and selects either the first response voice or the second response voice according to the comparison result. apparatus. The spoken dialogue apparatus according to claim 1, wherein the voice analysis unit sets a representative value of the prosodic index value in a plurality of past utterances as the threshold value. The response generation unit selects the first response sound when the prosodic index value is a numerical value outside a predetermined range including the threshold value, and when the prosodic index value is a numerical value inside the predetermined range, the second response The voice interactive apparatus according to any one of claims 1 to 3, wherein a voice is selected. The voice interaction apparatus according to claim 1, wherein the response generation unit reproduces the first response voice with respect to an utterance voice randomly selected from a plurality of utterance voices. The voice interaction apparatus according to claim 5, wherein the response generation unit variably sets the reproduction frequency of the first response voice with respect to the plurality of uttered voices. The voice interaction device according to claim 6, wherein the response generation unit sets the reproduction frequency of the first response voice according to a voice conversation usage history. Computer
An audio acquisition unit for acquiring an utterance signal representing the utterance audio; and
A program that causes a playback device to function as a response generation unit that selectively plays back a first response voice that represents a response to a spoken voice and a second response voice other than the response.

Images