JP4758931B2 - Speech synthesis apparatus, method, program, and recording medium thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide technology which alleviates sense of incongruity felt by a user when speech is changed from operator's speech to synthesized speech in a speech guidance system. <P>SOLUTION: A speaker identifying section 42 uses a feature quantity which is extracted from the speech of the operator 16 to select the most similar identification model to speech of the operator 16 from among a plurality of identification models prepared beforehand. A synthesized speech creation section 43 reads a speech data corresponding to the selected identification model, and creates synthesized speech close to quality of operator's speech, which corresponds to information regarding speech provided for the user determined by the operator. <P>COPYRIGHT: (C)2009,JPO&amp;INPIT

Description

The present invention relates to an apparatus, method, program, and recording medium for synthesizing speech .

  Conventionally, there has been a voice guidance system as described below (for example, see Non-Patent Document 1).

  A user interacts with a call center operator via a communication network. The operator selects an answer voice suitable for the user from the answer voice candidates registered in advance through the dialogue. The voice synthesizer reproduces the selected answer voice instead of the operator. After selecting the answer voice, the operator is released from the user, interacts with other users, and selects an appropriate answer voice in the same manner.

  As described above, the operator can only interact with the user, select the answer voice, and cause the voice synthesizer to play the answer voice, thereby reducing the response time of the operator per user. In other words, since the number of cases that can be handled per operator can be increased, the number of operators per number of cases that can be handled can be reduced. As a result, the voice guidance system according to the prior art can reduce the cost of the call center.

However, in the voice guidance system according to the prior art, the voice reproduced by the voice synthesizer is created without considering the voice of the operator who has interacted with the user. Therefore, the user has listened to the announcement by the synthesized voice having different speaker characteristics from the voice of the operator who has been in conversation until immediately before.
NTT IT Corporation, [online], November 30, 2004, [March 16, 2007 search], Internet <URL: http://www.ntt-it.co.jp/press/2004/ 041130 / 041130vcj.html>

  In the voice guidance system according to the prior art, the talker nature of the voice of the operator who had the initial conversation is different from the talker nature of the answer voice reproduced by the speech synthesizer, which makes the user feel unnatural or uncomfortable. There was a problem.

  According to the first aspect of the present invention, in the voice synthesizer in the voice guidance system that generates the synthesized voice by the voice synthesizer and provides the user with the synthesized voice instead of the call center operator, the voice synthesizer Speaker identification means and synthesized speech generation means, wherein the speaker identification means is an identification model storage means for storing a plurality of identification models, and a first feature quantity for extracting a first feature quantity from the input operator's voice The extraction means, the first feature amount is input to each identification model read from the identification model storage means, the likelihood calculation means for calculating the likelihood for each identification model, and the calculated likelihood is maximized Selection means for selecting an identification model to be selected from a plurality of identification models prepared in advance, and the synthesized speech generation means is a sound for storing speech data corresponding to the plurality of identification models. The data storage means, the voice data corresponding to the selected identification model are provided to the user determined by the input operator using the voice data selection means for reading from the voice data storage means and the read voice data. Synthesizing means for generating synthesized speech having a quality close to the quality of the operator's speech corresponding to the information related to the speech, and the speech synthesizer sets a speaker characteristic parameter representing the individuality of the speaker from the operator's speech. A second feature amount extracting means for extracting, a deforming means for generating a deformed speech by transforming the synthesized speech so that the sound quality of the synthesized speech approximates the sound quality of the operator's speech using the speaker characteristics parameter, and the calculated likelihood If the maximum value of the degree is greater than or greater than the predetermined threshold value α, or the calculated maximum likelihood value is smaller than the predetermined threshold value β that is a threshold value that is larger than the threshold value α. If not more than, and outputs a modified audio, otherwise, the speech synthesis apparatus for and control means for controlling to output the synthesized speech, and further comprising a.
  According to the second aspect of the present invention, in the voice synthesizer in the voice guidance system in which the voice synthesizer generates synthesized voice and provides the user with the synthesized voice instead of the call center operator. Speaker identification means and synthesized speech generation means, wherein the speaker identification means is an identification model storage means for storing a plurality of identification models, and a first feature quantity for extracting a first feature quantity from the input operator's voice The extraction means, the first feature amount is input to each identification model read from the identification model storage means, the likelihood calculation means for calculating the likelihood for each identification model, and the calculated likelihood is maximized Selection means for selecting an identification model to be selected from a plurality of identification models prepared in advance, and the synthesized speech generation means is a sound for storing speech data corresponding to the plurality of identification models. The data storage means, the voice data corresponding to the selected identification model are provided to the user determined by the input operator using the voice data selection means for reading from the voice data storage means and the read voice data. Synthesizing means for generating synthesized speech having a quality close to the quality of the operator's speech corresponding to the information related to the speech, and the speech synthesizer sets a speaker characteristic parameter representing the individuality of the speaker from the operator's speech. A second feature amount extracting means for extracting, a deforming means for generating a deformed speech by transforming the synthesized speech so that the sound quality of the synthesized speech approximates the sound quality of the operator's speech using the speaker characteristics parameter, and the calculated likelihood If the maximum value of the degree is greater than or greater than a predetermined threshold α and smaller than or less than a predetermined threshold β that is a threshold greater than the threshold α, the modified voice Outputs, otherwise, the speech synthesis device comprising control means for controlling to output the synthesized speech, further comprising a.

  By making the speech nature of the speech reproduced by the speech synthesizer closer to the speech nature of the voice of the operator who had the initial conversation, the user is not given unnaturalness or discomfort. Thereby, a user's satisfaction degree can be raised.

[First embodiment]
A functional configuration of the voice guidance system 1 according to the first embodiment will be described. As shown in FIG. 1, the voice guidance system 1 is connected to an input / output unit 17 for the user 10 to perform a dialog and an input / output unit 15 for the operator 16 to perform a dialog via a communication network 13. ing. The communication network 13 is an arbitrary communication network such as a telephone network, a cellular phone network, an IP communication network, and the Internet, and may be a wired / wireless, a public network, or a dedicated line as long as it can perform a conversation. The input / output units 15 and 17 are a telephone handset, a mobile phone, a handset, and the like.

  In response to an instruction from the operator 16 between the communication network 13 and the input / output unit 15, the switching unit 12 switches the connection destination of the input / output unit 17 to either the input / output unit 15 or the speech synthesizer 11. Is provided. In the initial state, the switching unit 12 is connected to the input / output unit 15.

  Further, for example, a first acquisition unit 14a that acquires information related to the voice provided to the user determined by the operator 16 and a second acquisition unit 14b that acquires the voice of the operator 16 are provided. 14 a passes the acquired information, and the second acquisition unit 14 b passes the acquired operator voice to the voice synthesizer 11. The first acquisition unit 14a is an input unit such as a keyboard, a pointing device, or a microphone. The second acquisition unit 14b is input means such as a microphone. The speech synthesizer 11 uses the received information and the voice of the operator 16 to generate a synthesized voice having a talkability close to that of the voice of the operator 16 and provides it to the user 10. .

  With reference to FIG. 7, the flow of processing of the voice guidance system 1 will be described. FIG. 7 is a flowchart illustrating the processing flow of the voice guidance system 1.

<Step S1>
The user 10 interacts with the operator 16 via the communication network 13 using the input / output unit 17. In the dialogue, the user 10 conveys the characteristics of the information he / she wants to obtain. The operator 16 specifies information that the user 10 wants to obtain.

<Step S2>
The operator 16 determines the voice to be provided to the user through the dialogue. The operator 16 inputs information related to the voice to be provided to the determined user to the first acquisition unit 14a. Thereby, the information regarding the voice provided to the user is acquired by the first acquisition unit 14 a and is output to the voice synthesizer 11. The determined information is, for example, information such as text information and identification information.

  For example, when the user is inquiring the telephone number of the operator, the determined information is text information composed of 10 numbers such as “03-xxxx-xxxx”. . X is a number from 0 to 9.

  Moreover, the voice provided to the user 10 is prepared as text information, and different identification information is determined in advance for each text information. Then, the operator 16 may determine the identification information and input it to the first acquisition unit 14a. In this case, this identification information is passed to the speech synthesizer 11. In this case, the speech synthesizer 11 stores text information corresponding to each identification information in a storage unit (not shown). Then, text information corresponding to the passed identification information is read from the storage unit and a speech synthesis process described later is performed.

<Step S3>
Switching command information for switching the switching unit 12 is input by an operator 16 from input means (not shown). The switching command information is passed to the switching unit 12. The switching unit 12 receives the switching command information, and switches the switching unit 12 so that the sound output from the speech synthesizer 11 is output to the user 10. That is, the voice that the user 10 listens to is switched from the voice of the operator to the voice output by the voice synthesizer 11.

<Step S4>
The voice of the operator 16 is input to the second acquisition unit 14 b and passed to the voice synthesizer 11. Here, the voice of the operator 16 may include not only the voice of the operator 16 but also a sound generated from the environment around the operator 16 collected by the second acquisition unit 14b or the input / output unit 15. .

<Step S5>
The voice synthesizer 11 uses the input information and the operator's voice to generate a synthesized voice close to the sound quality of the operator's voice and provides it to the user 10.

  Hereinafter, the functional configuration of the speech synthesizer 11 and details of the processing in step S5 will be described with reference to FIGS. FIG. 2 is a diagram illustrating the functional configuration of the speech synthesizer 11, and FIG. 8 is a flowchart illustrating the flow of each process constituting step S <b> 5. Since FIG. 2 is a diagram for explaining the speech synthesizer 11 with emphasis, it should be noted that the input / output unit 15, the switching unit 12, the communication network 13, and the like are omitted.

  As illustrated in FIG. 2, the speech synthesizer 11 uses the input operator's voice to select a speaker identification unit 42 that selects a suitable identification model from among identification models prepared in advance, and the operator selects And a synthesized speech generation unit 43 that receives the information about the identification model selected by the speaker identification unit 42 (for example, the ID of the identification model) and generates a synthesized speech that is close to the speech characteristics of the operator's speech. Is provided.

  The speaker identification unit 42 includes a first feature amount extraction unit 421 that extracts a first feature amount from the operator's voice, and an identification model storage unit that stores each of the identification models 4221 to 422n that are prepared in advance and assigned different IDs. 422, a likelihood calculation unit 420 that inputs the extracted feature quantity to each identification model read from the identification model storage unit 422, calculates a likelihood for each identification model, and outputs the likelihood to the selection unit 423; For example, a selection unit 423 that selects an ID of an identification model that maximizes the calculated likelihood and outputs the ID to the synthesized speech generation unit 43 is provided.

  The synthesized speech generation unit 43 stores speech data corresponding to each identification model 4221 to 422n, speech data storage unit 433 storing speech data 4331 to 433n, and speech data corresponding to the ID of the identification model selected by the selection unit 423. Using the voice data selection unit 432 that is read from the voice data storage unit 433 and output to the synthesis unit 431, the information acquired by the first acquisition unit 14a, and the voice data selected by the voice data selection unit 432, an operator is used. 16 includes, for example, a synthesis unit 431 that generates synthesized speech having sound quality close to that of the operator's speech, corresponding to information related to speech provided to the user 10 selected by 16.

<< Step S51 (FIG. 8) >>
The first feature quantity extraction unit 421 extracts the feature quantity from the input voice of the operator 16 and sends it to the likelihood calculation unit 420. As the feature amount, for example, an arbitrary acoustic feature amount or speech feature amount such as a spectrum envelope or MFCC (Mel-Frequency Cepstrum Coefficient) can be used. Moreover, the calculation method of those feature-values is also arbitrary. In the present embodiment, the case where MFCC generally used in speech recognition, speaker identification, and the like is used as the feature amount will be described as an example. When MFCC is used as the feature amount, the MFCC is calculated for each frame t obtained by dividing the voice by a predetermined time length.

<< Step S52 >>
The likelihood calculation unit 420 calculates the likelihood for each identification model using the feature amount calculated by the first feature amount extraction unit 421 and the identification model read from the identification model storage unit 422, and selects the selection unit. 423. In this embodiment, together with the likelihood for each identification model, the ID assigned to that identification model is also sent.

  The process of the likelihood calculation unit 420 will be described by taking as an example a case where the identification model is expressed using a mixed normal distribution (hereinafter referred to as GMM (Gaussian Mixture Model)).

The MFCC of frame t is X _t , the number of mixture distributions is M, the model parameter is λ, the d-dimensional Gaussian function is b _i (•), the mixture weight is c _i (Σ _i c _i = 1), and U _i is set. If the mean vector, σ _i is a covariance matrix, and • ^T is a transpose of •, the first likelihood P (X _t | λ) in the frame t is expressed as follows.

The model parameter λ is composed of M model parameters λ _i (i = 1,..., M), and each model parameter λ _i is obtained from the mixture weight c _i , the average vector U _i , and the covariance matrix σ _i. Composed.

The mixture weight c _i , average vector U _i , and covariance matrix σ _i constituting the model parameters are calculated in advance from the speech data corresponding to each identification model. That is, the model parameter λ is calculated for each audio data. The model parameters λ calculated for each of these identification models are given different IDs and stored in advance in the identification model storage unit 422. Reading the identification model means reading the corresponding model parameter λ.

In the present embodiment, the likelihood calculation unit 420 has a first likelihood P (X _t | λ) for each of a plurality of frames (for example, each frame from the first frame 1 to the T-th frame T). And take its logarithm. Then, by adding them, the second likelihood P (X | λ) is calculated and sent to the selection unit 423.

<< Step S53 >>
The selection unit 423 selects an identification model that maximizes the likelihood (in this embodiment, the second likelihood P (X | λ)) calculated by the likelihood calculation unit 420 for each identification model. Then, information about the identification model to be maximized is sent to the synthesized speech generation unit 43. Here, the information about the identification model is an ID given to the audio data corresponding to the identification model.

<< Step S54 >>
The voice data selection unit 432 of the synthesized voice generation unit 43 reads voice data corresponding to the identification model selected by the selection unit 423 from the voice data storage unit 433 and sends the voice data to the synthesis unit 431. Specifically, the synthesized voice generation unit 43 reads voice data having the same ID as the ID of the identification model sent from the selection unit 423 from the voice data storage unit 433.

<< Step S55 >>
The synthesis unit 431 uses the voice data sent from the voice data selection unit 432 to generate synthesized voice having a sound quality close to the voice quality of the operator's voice corresponding to the information related to the voice provided to the user determined by the operator. . The generated synthesized speech is provided to the user 10.

  Any known technique may be used as a method of creating the synthesized speech. In the following, a case where waveform connection type speech synthesis is used as the synthesis method and text information is used as information about the speech provided to the user 10 determined by the operator 16 will be described as an example.

  As illustrated in FIG. 3, the synthesis unit 431 includes a text analysis unit 4311, a prosody generation unit 4312, a synthesis processing unit 4313, a reading / accent dictionary storage unit 4314, and a prosody generation rule storage unit 4315.

  With reference to FIG. 9, a processing example of the synthesis unit 431 will be described. FIG. 9 is a diagram illustrating an example of the processing flow of the synthesis unit 431.

<< Step S551 >>
The text analysis unit 4311 performs morphological analysis on the input text information, refers to the reading / accent dictionary read from the reading / accent dictionary storage unit 4314 from the obtained morpheme information, and reads the reading / accent of the input text information. decide. Information on the analyzed morpheme, the determined reading, and the determined accent is sent to the prosody generation unit 4312 and the synthesis processing unit 4313.

For example, if the text information "Do you want to call?" Is input, text analysis section 4311, performs a morphological analysis as "Odenwa [02] Shimasuka? [02]", to determine the reading and accent. Here, the underlined portion is a reading portion, and [02] represents an accent portion.

<< Step S552 >>
The prosody generation unit 4312 generates a prosody of the synthesized speech by referring to the prosody generation rule read from the prosody generation rule storage unit 4315 from the analyzed morpheme, the determined reading, and the determined accent. The generated prosody is sent to the synthesis processing unit 4313.

For example, the prosody can be represented by a curve in a graph with the horizontal axis representing time and the vertical axis representing the fundamental frequency, as illustrated on the right side of the prosody generation unit 4312 in FIG. Curve in the graph illustrated on the right side of the prosody generation part 4312 of FIG. 3 is a prosody which corresponds to the information related to "Odenwa [02] Shimasuka? [02]".

<< Step S553 >>
The synthesis processing unit 4313 selects and connects speech that matches the analyzed morpheme, the determined reading, the determined accent, the determined prosody, and the like from the speech data 4316 to generate a synthesized speech. . The voice data 4316 is the voice data sent from the voice data selection unit 432, that is, one voice selected by the voice data selection unit 432 from the voice data storage unit 433 corresponding to the identification model selected by the selection unit 423. It is data. As illustrated in FIG. 3, the audio data 4316 stores a plurality of audio data in which the phoneme of the corresponding speech, the phonemes before and after the corresponding speech, and the prosody are associated with each other. The synthesis processing unit 4313 refers to the audio data 4316 using the phoneme of the corresponding speech, the phonemes before and after the corresponding speech, and the prosody as an index, and selects and connects the speech from the speech data 4316 to generate a synthesized speech ( See, for example, Reference 1.)

[Reference 1] Toshio Hirai, 7 others, “Corpus-based Speech Synthesis System XIMERA”, IEICE Technical Report, The Institute of Electronics, Information and Communication Engineers, SP2005-18 (2005-5), p. 37-42
In this way, by generating synthesized speech from speech data corresponding to the identification model that maximizes the likelihood, synthesized speech that is close to the voice quality of the operator 16 voice or voice quality, in other words, that reflects the operator's speech characteristics Voice can be generated. Thereby, it is possible to reduce a sense of incongruity or a gap when the voice that the user 10 listens to is switched from the voice of the operator 16 to the synthesized voice.

[Second Embodiment]
The voice guidance system according to the second embodiment includes a voice synthesizer 21 illustrated in FIG. 4 instead of the voice synthesizer 11, and the voice synthesizer 21 deforms the generated synthesized voice to further change the operator 16. It differs from the voice guidance system 1 according to the first embodiment in that a modified voice close to the voice quality is generated. That is, the voice synthesizer 21 is different from the voice guidance system according to the first embodiment in that the process of step S56a indicated by the broken line in FIG. 8 is performed. Since other functional configurations and processes are the same as those in the first embodiment, the description thereof is omitted. FIG. 4 is a diagram illustrating a functional configuration of the speech synthesizer 21. FIG. 10 is a flowchart illustrating the process flow of step S56a.

  As illustrated in FIG. 4, the speech synthesizer 21 includes a speaker identification unit 42, a synthesized speech generation unit 43, a second feature amount extraction unit 64, and a deformation unit 65. The speaker identification unit 42 and the synthesized speech generation unit 43 are the same as those in the first embodiment.

<< Step S56a1 >>
The second feature amount extraction unit 64 extracts a second feature amount (speaker parameter) from the operator's voice and sends it to the deformation unit 65. The speaker characteristic parameter, which is the second feature amount, is a parameter representing the individuality of the voice speaker (speaker). For example, an arbitrary acoustic feature quantity such as a fundamental frequency, an utterance speed, and a power can be used as the speaker property parameter. Any known technique may be used as a method for extracting the speaker characteristic parameter. For example, the second feature amount extraction unit 64 divides the operator's voice into frames of a certain time length, extracts the fundamental frequency for each frame, and sends the basic frequency to the deformation unit 65.

<< Step S56a2 >>
The deforming unit 65 uses the speaker property parameter to deform the synthesized speech so that the quality of the synthesized speech generated by the synthesized speech generating unit 43 approaches the quality of the voice of the operator 16 to generate a modified speech. The generated modified voice is provided to the user 10. Any known technique may be used as a method of transforming sound.

  Hereinafter, the case where the fundamental frequency is used as the speaker property parameter will be described as an example. In this case, the deforming unit 65 includes a fundamental frequency extracting unit 651, an average value calculating unit 652, a parallel moving unit 653, and a recombining unit 654, as illustrated in FIG.

  The fundamental frequency extraction unit 651 extracts the fundamental frequency of the synthesized speech generated by the synthesized speech generation unit 43 and sends it to the average value calculation unit 652. Specifically, the fundamental frequency extraction unit 651 divides the synthesized speech into frames having a certain time length, and extracts the fundamental frequency for each frame. As a method for obtaining the fundamental frequency, any method such as a cepstrum method, a lag method, or a TEMPO method can be used.

  The average value calculation unit 652 obtains the average value of the fundamental frequency extracted by the fundamental frequency extraction unit 651 and the average value of the fundamental frequency extracted by the second feature amount extraction unit 64, and sends them to the translation unit 653. Here, the average value of the fundamental frequency is, for example, the average value of the fundamental frequency in the voice that the operator reads out a sentence of a predetermined length, the average value of the fundamental frequency in a certain time length, and the fundamental frequency in a certain number of frames. Is the average value. As the one sentence, the time length, and the number of frames are longer or larger, the average value of the fundamental frequency can be obtained with higher accuracy. That is, it becomes easy to extract the height and lowness of the voice unique to the speaker, and the effect of the present embodiment is increased accordingly.

  The translation unit 653 adjusts the fundamental frequency so that the average value of the fundamental frequency extracted by the fundamental frequency extraction unit 651 is matched with the average value of the fundamental frequency extracted by the second feature amount extraction unit 64. The outline (similar to the curve illustrated on the right side of the prosody generation unit 4312 in FIG. 3) is changed by translation.

  The re-synthesizing unit 654 re-synthesizes the voice in accordance with the changed basic frequency to generate a modified voice. As a method for re-synthesis, a signal processing method such as PSOLA can be used (for example, see Reference 2).

[Reference 2] Ryo Mochizuki and Tetsunori Kobayashi, “A low-band spectrum envelope modeling for high quality pitch modification”, Proc. IEEE International Conference on Acoustics, Speech, and Signal, Processing (ICASSP 2004), Vol.1, pp .645-648 (2004-5).
In this way, by using the second feature amount extracted from the voice of the operator 16 and modifying the synthesized voice so as to approach the voice quality of the voice of the operator 16, a modified voice closer to the voice quality of the voice of the operator 16. Can be generated. Thereby, it is possible to further reduce a sense of incongruity and a gap when switching voices.

  Further, according to the present embodiment, it is possible to generate a modified speech that sufficiently reflects the operator's speaker characteristics without preparing a set of an identification model and speech data sufficiently close to the operator's speech. For this reason, even if the number of sets of the identification model and voice data prepared in advance is reduced, there is an advantage that it is possible to reduce a sense of incongruity and a gap when switching voices.

[Third embodiment]
The voice guidance system according to the third embodiment includes a voice synthesizer 31 illustrated in FIG. 6 instead of the voice synthesizer 21, and the voice synthesizer 31 is a modification for making the synthesized voice similar to the voice of the operator. It is at least different from the voice guidance system according to the second embodiment in that control is performed such that the deformation is not performed when the amount increases. That is, the voice synthesizer 21 is different from the voice guidance system according to the second embodiment in that the process of step S56b indicated by the alternate long and short dash line in FIG. 8 is performed. Other functional configurations and processes are basically the same as those in the second embodiment, and thus the description thereof is omitted. FIG. 6 is a diagram illustrating a functional configuration of the speech synthesizer 31. FIG. 11 is a flowchart illustrating the process flow of step S56b.

  As illustrated in FIG. 6, the speech synthesizer 31 includes, for example, a speaker identification unit 42, a synthesized speech generation unit 43, a second feature amount extraction unit 64, a deformation unit 65, and a control unit 80. The control unit 80 includes, for example, a determination unit 83 that determines which one of synthesized speech and modified speech is to be generated and output, and switching units 81a to 81c that switch each connection according to the determination result of the determination unit 83. .

<< Step S56b1 (FIG. 8) >>
The determination unit 83 of the control unit 80 determines whether or not the maximum likelihood value is equal to or greater than the predetermined threshold value α and equal to or smaller than the predetermined threshold value β, that is, whether α ≦ maximum likelihood value ≦ β. Judging. The maximum likelihood is the likelihood of the identification model that maximizes the likelihood selected by the selection unit 423 (FIG. 2) of the speaker identification unit 42. In the third embodiment, the maximum likelihood value is passed from the speaker identification unit 42 to the determination unit 83.

  In this example, the determination unit 83 determines whether α ≦ maximum value of likelihood ≦ β, but instead α <maximum value of likelihood <β, α <maximum value of likelihood ≦ It may be determined whether β, α ≦ maximum likelihood <β.

  The predetermined threshold value α and the predetermined threshold value β vary depending on the definition of likelihood, the method of generating synthesized speech, the method of generating modified speech, the quality of required speech, and the like, so that they are appropriately determined so that appropriate results can be obtained.

  The reason why the control unit 80 uses a criterion as to whether or not α ≦ maximum likelihood ≦ β will be described later.

<< Step S56b2 >>
If α ≦ maximum value of likelihood ≦ β, the control unit 80 performs control so as to deform the synthesized speech to generate and output a modified speech. Specifically, the determination unit 83 sends a signal for generating and outputting the modified sound to each of the switching units 81a to 81c, and the switching units 81a to 81c switch the switches in response to the signal. In this example, each switching part 81a-81c switches a switch so that it may connect with the terminal by which Yes of FIG. 6 was displayed.

  Then, the speech synthesizer 31 performs processing of step S56b3 (generation of synthesized speech), step S56b4 (extraction of the second acoustic feature amount), and step S56b5 (generation of modified speech) of FIG. The processing of step S56b3 and step S56b4 is the same as the processing of step S55 (FIG. 8), step S56a1 (FIG. 10), and step S56a2 described in the first embodiment and the second embodiment, and thus description thereof is omitted.

<< Step S56b6 >>
If α ≦ maximum likelihood value ≦ β is not satisfied, the control unit 80 performs control so as to output synthesized speech without generating modified speech. Specifically, the determination unit 83 sends a signal indicating that the modified sound is not generated / output to the switching units 81a to 81c, and the switching units 81a to 81c switch the switches in response to the signal. In this example, each switching part 81a-81c switches a switch so that it may connect with the terminal by which No of FIG. 6 was displayed.

  Then, the speech synthesizer 31 performs step S56b7 (generation of synthesized speech) in FIG. Since the process of step S56b7 is the same as that of step S55 (FIG. 8) demonstrated in 1st embodiment, description is abbreviate | omitted.

  When the maximum likelihood is equal to or greater than the predetermined threshold α, the quality of the synthesized speech is already sufficiently close to the quality of the operator's speech. For this reason, even if the modified speech is generated from the synthesized speech, the quality of the generated modified speech may not be close to the quality of the operator's speech, and may be further away from the quality of the operator's speech.

  On the other hand, when the maximum likelihood value is less than or equal to the predetermined threshold value β, the amount of deformation to approximate the sound quality of the operator's voice is large. For this reason, when the modified speech is generated from the synthesized speech, the generated modified speech may be close to the sound quality of the operator's speech, but may be of low quality with reduced naturalness and intelligibility. is there.

  For this reason, in the present embodiment, the control unit 80 determines whether to generate / output synthesized speech or generate / output modified speech, based on whether or not α ≦ maximum value of likelihood ≦ β. I have decided.

  In this way, if the amount of deformation when generating modified speech from synthesized speech increases and the naturalness and intelligibility may deteriorate, the synthesized speech can be generated without generating / outputting the modified speech. By performing the control of output, it is possible to achieve both a sense of incongruity / gap due to voice switching and securing the naturalness / intelligibility of the voice provided to the user 10.

[Modifications, etc.]
The order of the processes in steps S2 to S4 illustrated in FIG. 7 is arbitrary. That is, any process of steps S2 to S4 may be performed first, and the processes of these steps may be performed simultaneously.

  Instead of the two acquisition means of the first acquisition unit 14a (FIG. 1) and the second acquisition unit 14b, a third acquisition unit 14c that is one acquisition unit including an electroacoustic transducer such as a microphone may be provided. In this case, the operator 16 inputs information to be provided to the user into the third acquisition unit 14c with his / her own voice. Information input by the voice of the operator 16 is output to the speech synthesizer 11.

  Instead of performing the process of step S4 (FIG. 7), as indicated by a broken line in FIG. 1, in the dialogue between the user 10 and the operator 16 in step S1, the voice of the operator 16 collected by the input / output unit 15 May be input to the speech synthesizer 11.

  As indicated by a broken line in FIG. 11, the process of step S560 for generating synthesized speech may be performed before step S56b1. In this case, the process of step S56b3 and step S56b7 is not performed. Other processes are the same as described above.

  In step S56b1 (FIG. 11), the determination unit 83 (FIG. 6) determines that α ≦ maximum likelihood value, α <maximum likelihood value, maximum likelihood value ≦ β, and maximum likelihood value <β. One determination may be made to determine whether to generate synthesized speech or to generate modified speech.

  In step S56b6 (FIG. 11), when α ≦ maximum likelihood value ≦ β is not satisfied, the switch may be switched so that only the switching unit 81a (FIG. 6) and the switching unit 81c are connected to the No terminal. . That is, the control unit 80 may perform control so as to generate modified speech as long as control is performed so as to finally output synthesized speech.

  When the speech synthesizers 11, 21, and 31 are realized by a computer, processing contents of functions that each device should have are described by a program. Then, by executing this program on the computer illustrated in FIG. 12, the above processing functions are realized on the computer.

The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, the magnetic recording device may be a hard disk device or a flexible Discs, magnetic tapes, etc. as optical discs, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD
-R (Recordable) / RW (ReWritable), etc., MO (Magneto-Optical disc), etc. as a magneto-optical recording medium, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. as a semiconductor memory it can.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  As an execution form different from the above-described embodiment, the computer may read the program directly from the portable recording medium and execute processing according to the program. Each time is transferred, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, each apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.

The figure which illustrates the function structure of the voice guidance system. The figure which illustrates the functional structure of the speech synthesizer 11 by 1st embodiment. The figure which illustrates the function structure of the synthetic | combination part 431. FIG. The figure which illustrates functional composition of speech synthesizing device 21 by a second embodiment. The figure which illustrates the function structure of the deformation | transformation part. The figure which illustrates functional composition of speech synthesizing device 31 by a third embodiment. The figure which illustrates the flow of a process of the voice guidance system. The figure which illustrates the flow of processing of Step S5. The figure which illustrates the flow of processing of Step S55. The figure which illustrates the flow of processing of Step S56a. The figure which illustrates the flow of processing of Step S56b. The figure which illustrates the function structure in the case of implement | achieving a speech synthesizer with a computer.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Voice guidance system 10 User 11 Voice synthesizer 12 Switching part 13 Communication network 14a First acquisition part 14b Second acquisition part 14c Third acquisition part 15 Input / output part 16 Operator 17 Input / output part 21 Voice synthesizer 31 Voice synthesizer 42 Speaker identification unit 43 Synthetic speech generation unit 64 Second feature amount extraction unit 65 Deformation unit 80 Control unit 83 Determination unit 420 Likelihood calculation unit 421 First feature amount extraction unit 422 Identification model storage unit 423 Selection unit 431 Synthesis unit 432 Speech Data selection unit 433 Audio data storage unit 422n Identification model 433n Audio data

Claims (6)

In a speech synthesizer in a speech guidance system that generates synthesized speech by a speech synthesizer and provides information to the user with synthesized speech instead of a call center operator,
The speech synthesizer includes a speaker identification unit and a synthesized speech generation unit.
The speaker identification means is
An identification model storage means for storing a plurality of identification models;
From the speech of the inputted said operator and first feature extraction means for extracting a first feature quantity,
The first feature amount, and input to the identification model read from the identification model storage unit, a likelihood calculation means for calculating a likelihood for each identified model,
Selecting means for selecting an identification model that maximizes the calculated likelihood from a plurality of identification models prepared in advance ;
With
The synthetic speech generating means is
Voice data storage means for storing voice data corresponding to each of the plurality of identification models;
Voice data selection means for reading voice data corresponding to the selected identification model from the voice data storage means;
Using the speech data read out above, it corresponds to the information about the audio to be provided to the user that entered the operator has determined, synthesizing means for generating a synthetic speech having a quality close to the quality of the voice of the operator ,
Bei to give a,
The above speech synthesizer
A second feature amount extracting means for extracting a speaker characteristic parameter representing the individuality of the speaker from the voice of the operator;
Deformation means for deforming the synthesized speech and generating a modified speech so that the quality of the synthesized speech approaches the quality of the operator's speech using the speaker property parameter ;
If the maximum value of the calculated likelihood is greater than or greater than a predetermined threshold α, or the maximum value of the calculated likelihood is smaller than a predetermined threshold β that is a threshold larger than the threshold α or if less, then it outputs the modified speech, otherwise, control means for controlling to output the synthesized speech,
A speech synthesizer further comprising: In a speech synthesizer in a speech guidance system that generates synthesized speech by a speech synthesizer and provides information to the user with synthesized speech instead of a call center operator,
The speech synthesizer includes a speaker identification unit and a synthesized speech generation unit.
The speaker identification means is
An identification model storage means for storing a plurality of identification models;
First feature quantity extraction means for extracting a first feature quantity from the input voice of the operator;
The first feature amount, and input to the identification model read from the identification model storage unit, a likelihood calculation means for calculating a likelihood for each identified model,
Selecting means for selecting an identification model that maximizes the calculated likelihood from a plurality of identification models prepared in advance ;
With
The synthetic speech generating means is
Voice data storage means for storing voice data corresponding to each of the plurality of identification models;
Voice data selection means for reading voice data corresponding to the selected identification model from the voice data storage means;
Using the speech data read out above, it corresponds to the information about the audio to be provided to the user that entered the operator has determined, synthesizing means for generating a synthetic speech having a quality close to the quality of the voice of the operator ,
Bei to give a,
The above speech synthesizer
A second feature amount extracting means for extracting a speaker characteristic parameter representing the individuality of the speaker from the voice of the operator;
Deformation means for deforming the synthesized speech and generating a modified speech so that the quality of the synthesized speech approaches the quality of the operator's speech using the speaker property parameter ;
If the calculated maximum likelihood is greater than or greater than a predetermined threshold α and smaller than or less than a predetermined threshold β that is a threshold greater than the threshold α , the modified speech is output. and, otherwise, control means for controlling to output the synthesized speech,
A speech synthesizer further comprising:   In a voice synthesis method in a voice guidance method for generating synthesized voice by a voice synthesis method and providing information to a user with synthesized voice instead of a call center operator,
  The speech synthesis method includes a speaker identification step and a synthesized speech generation step,
  The speaker identification step is
  A first feature amount extraction step for extracting a first feature amount from the input voice of the operator;
  A likelihood calculation step of inputting the first feature amount into each identification model read from an identification model storage means for storing a plurality of identification models, and calculating a likelihood for each identification model;
  A selection step of selecting an identification model that maximizes the calculated likelihood from a plurality of identification models prepared in advance;
  Have
  The synthetic speech generation step includes
  A voice data selection step of reading voice data corresponding to the selected identification model from voice data storage means for storing voice data corresponding to each of the plurality of identification models;
  Using the read voice data to generate a synthesized voice having a sound quality close to that of the operator's voice corresponding to the information related to the voice provided to the user determined by the operator; ,
  Have
  The above speech synthesis method
  A second feature extraction step for extracting a speaker characteristic parameter representing the individuality of the speaker from the operator's voice;
  Using the speaker property parameter, a deformation step for deforming the synthesized speech to generate a modified speech so that the sound quality of the synthesized speech approaches the sound quality of the operator's speech;
  If the maximum value of the calculated likelihood is greater than or greater than a predetermined threshold α, or the maximum value of the calculated likelihood is smaller than a predetermined threshold β that is a threshold larger than the threshold α or A control step for controlling to output the modified voice if not, and to output the synthesized voice otherwise;
  The speech synthesis method further comprising:   In a voice synthesis method in a voice guidance method for generating synthesized voice by a voice synthesis method and providing information to a user with synthesized voice instead of a call center operator,
  The speech synthesis method includes a speaker identification step and a synthesized speech generation step,
  The speaker identification step is
  A first feature amount extraction step for extracting a first feature amount from the input voice of the operator;
  A likelihood calculation step of inputting the first feature amount into each identification model read from an identification model storage means for storing a plurality of identification models, and calculating a likelihood for each identification model;
  A selection step of selecting an identification model that maximizes the calculated likelihood from a plurality of identification models prepared in advance;
  Have
  The synthetic speech generation step includes
  A voice data selection step of reading voice data corresponding to the selected identification model from voice data storage means for storing voice data corresponding to each of the plurality of identification models;
  Using the read voice data to generate a synthesized voice having a sound quality close to that of the operator's voice corresponding to the information related to the voice provided to the user determined by the operator; ,
  Have
  The above speech synthesis method
  A second feature extraction step for extracting a speaker characteristic parameter representing the individuality of the speaker from the operator's voice;
  Using the speaker property parameter, a deformation step for deforming the synthesized speech to generate a modified speech so that the sound quality of the synthesized speech approaches the sound quality of the operator's speech;
  If the calculated maximum likelihood is greater than or greater than a predetermined threshold α and smaller than or less than a predetermined threshold β that is a threshold greater than the threshold α, the modified speech is output. Otherwise, a control step for controlling to output the synthesized speech;
  The speech synthesis method further comprising: Voice synthesizing program for causing a computer to function as the speech synthesis device according to claim 1 or 2. A computer-readable recording medium on which the speech synthesis program according to claim 5 is recorded.

Images