JP3881971B2 - Voice quality difference evaluation table creation device, voice corpus voice quality difference evaluation table creation system, and speech synthesis system - Google Patents

Description

  The present invention relates to a speech synthesis technology, and more particularly to a speech unit connection type speech synthesis technology that synthesizes speech that is close to a natural utterance by selecting and connecting speech units from a speech corpus.

  With the development of computer technology and data communication technology, the interface between humans and machines has become important. For human beings, it is desirable to communicate with machines in the same way as talking to people, and technical development for that purpose is underway.

  Cognitive techniques such as voice recognition and image recognition are mainly used for transmitting information from humans to machines. There are various methods for transmitting information from a machine to a human. Among them, there are increasing opportunities for using a speech synthesis technique, and a speech response system, a speech translation system, and the like are typical application examples. Furthermore, with the development of robots and the like in recent years, it is expected that communication between humans and robots can be realized in the same way as communication between humans by combining voice recognition and image recognition with voice synthesis.

  FIG. 8 shows a block diagram of a speech unit connection type speech synthesis system using a speech corpus. Referring to FIG. 8, in a speech unit connection type speech synthesis system using a speech corpus, a speech of a natural utterance by a human is recorded, and the speech segment of the utterance is corpusified in advance as speech corpus 40.

  When the input text 42 is given to this system, the speech unit selection unit 44 uses the speech unit corresponding to the speech constituting the input text from the speech corpus 40 to use the speech unit for speech synthesis. Select as a candidate. The speech element selection unit 44 evaluates the selected speech element by the evaluation function 50, and determines an optimal speech element to be used for speech synthesis according to the result. In this way, speech segments corresponding to the speech constituting the input text 42 are extracted. The speech unit connection unit 46 connects these series of speech units, so that synthesized speech data 48 corresponding to the input text 42 is generated.

  The evaluation function 50 is given, as a variable, a physical quantity that can be observed from the speech unit selection unit 44 and the speech unit selected before and the candidate speech unit. The evaluation function 50 outputs an evaluation value related to the given physical quantity as a dependent variable. Based on the evaluation value output by the evaluation function 50, the speech unit selection unit 44 determines a speech unit suitable for connection to the immediately preceding speech unit from the selected plurality of speech units.

  The voice synthesized in this way is synthesized using voice actually uttered by a human. Therefore, it is possible to synthesize a relatively natural voice that does not feel the so-called “mechanical sound”. Further, when the speech corpus becomes large-scale, the choices when the speech unit selection unit 44 selects speech units increase. Therefore, the speech unit selection unit 44 can determine a speech unit suitable for connection from among these many options, and the possibility of improving the quality of the synthesized speech is increased. For this reason, the current speech unit connection type speech synthesis system uses a large-scale speech corpus that records speech for several tens of hours.

  On the other hand, in a speech unit connection type speech synthesis system, continuous speech information is synthesized by connecting speech units extracted from speech recorded separately. Therefore, it is required that the voice quality of the connected speech elements is uniform. Otherwise, the synthesized speech has a discontinuity and the quality of the synthesized speech is degraded. Thus, in many cases, a speech corpus consisting of a single speaker's speech is used.

  In this case, since there is only one speaker, it is necessary to record speech over a long period of time in order to construct a large-scale speech corpus. Therefore, the utterance of one speaker is recorded in a plurality of recording periods (hereinafter, this recording period is referred to as “session”). In some cases, the recording may take months to years.

  As described above, even when a voice from a single speaker is recorded as described above, a recording period of several months may be required to create a large-scale voice corpus. During such a long recording period, it is extremely difficult to keep the sound recording conditions constant every day. In particular, it is extremely difficult for the speaker to keep his physical condition constant. For this reason, it is inevitable that the voice quality of the recorded audio data varies from session to session due to the change in the recording conditions. Therefore, there is a problem that the voice quality of the synthesized speech varies and a discontinuity is generated. Therefore, a technique for eliminating the unnaturalness when connecting speech waveform segments is desired.

  In order to eliminate unnaturalness when connecting speech segments, how to evaluate and select speech waveform segments used for synthesis becomes a problem. Usually, some acoustic feature quantity related to each speech waveform segment is calculated, and a speech waveform segment that matches a predetermined condition is selected. In order to reduce unnaturalness, it is important to perform segment selection using a scale that matches the perceptual characteristics as much as possible.

  In Non-Patent Document 1 described later, a cost is calculated for a candidate speech segment using an evaluation function called a “cost function” that reflects perceptual characteristics, and a waveform segment that minimizes the calculated cost is calculated. Selected. It is expected that more natural speech can be synthesized by selecting waveform segments using such a cost function.

  However, it is not clear what physical scale is used as the cost function to eliminate the unnatural perception at the time of waveform connection. That is, the correspondence between the physical scale and the naturalness of the synthesized speech is not clear. Therefore, in Non-Patent Document 1, the cost function is divided into a plurality of sub cost functions corresponding to various factors.

  When a corresponding physical quantity (observable) is given to each sub cost function, the sub cost is output as the function. By multiplying these sub-costs by weights and adding them, a cost as an evaluation value is calculated.

  In Non-Patent Document 1, a sub-cost function related to prosody, a sub-cost function related to discontinuity of F0 (fundamental frequency), a sub-cost function in phoneme environment substitution, a sub-cost function related to spectrum discontinuity, and a sub-cost function related to phoneme suitability are used. Of these sub-cost functions, the mapping between the perceptual evaluation and the physical quantity is performed especially regarding the phoneme environment substitution, which is a relatively easy to understand relationship with the perceptual evaluation. However, other factors do not use perceptual evaluation.

Tomoki Toda, Tsune Kawai, Minoru Tsuzaki, Kiyohiro Shikano, “Fragment Selection Based on Phoneme Units and Diphone Units in Segment-Connected Japanese Text Speech Synthesis”, IEICE Transactions, Vol. J85-D-II. , No. 12, pp. 1760-1770, Dec. 2002.

  In the technology described in Non-Patent Document 1, natural deterioration due to substitution of phonemic environment is evaluated by perceptual evaluation, and the result is reflected in the sub-cost function. However, non-patent document 1 does not consider the deterioration of the naturalness of the synthesized speech due to the voice quality of a speaker that is different for each session. This is because the correspondence between the voice quality difference due to the session difference and the physical characteristics of the voice is unknown or it is extremely difficult to identify it.

Therefore, an object of the present invention is to synthesize speech produced by speech quality changes caused by changes in the physical condition of a speaker when speech components extracted from speech data with different recording times are connected to synthesize speech. It is to provide a voice quality difference evaluation table creating device and a voice synthesizing system for reducing the deterioration of the voice quality of the voice .

Another object of the present invention is to provide a voice quality difference evaluation table creating apparatus and a voice synthesis system for evaluating a voice quality difference caused by a change in the physical condition of a speaker of voice data having different recording times with good sensitivity. That is.

  The voice quality difference evaluation table creation device according to the first aspect of the present invention is a device that creates a voice quality difference evaluation value table representing voice quality differences among a plurality of types of utterance voice data. This voice quality difference evaluation table creation device includes an extraction means for extracting an arbitrary combination of first and second utterance voice data from a plurality of types of utterance voice data, and an utterance of an arbitrary combination extracted by the extraction means Stimulus generation means for generating first and second audio stimuli based on each of the audio data, and an auditory sense for the first and second audio stimuli for each of the utterance audio data in any combination Perceptual test means for conducting a perceptual test on the voice quality difference between the speech utterances and deriving an evaluation value on the voice quality difference between the speech data of the combination, and the speech voice data of any combination based on the result of the perceptual test A table creation means for creating a voice quality difference evaluation value table by storing an evaluation value representing a difference in voice quality between the voice data and a combination of the speech data, No.

The difference voice between the speech data recorded by different recording conditions can be evaluated by sensory testing. Therefore, it is possible to evaluate the difference in voice quality of the uttered speech that is difficult to evaluate with a physical scale with good sensitivity.

Preferably, the perceptual test means includes a stimulus presenting means for comparing and presenting the first and second voice stimuli to the subject to be subjected to the perceptual test, and the first and second voice stimuli by the subject. Acquisition means for acquiring a result of evaluation relating to the magnitude of the difference in auditory voice quality between the two as a result of the perceptual test.

  When conducting a perceptual test, the subject clearly evaluates the difference in voice quality between the two voiced speech data to be compared by allowing the subject to compare the voice quality of the speech to be evaluated. can do.

  More preferably, the acquisition means has an evaluation value based on a means for obtaining a magnitude of the auditory difference rated by the subject using a predetermined rating scale and a value of the rating scale obtained by the rating by the subject. Evaluation value deriving means for deriving.

  By performing a perceptual test using the rating scale, a clearer result of the perceptual test can be obtained.

  The evaluation value deriving unit may include a unit for deriving an evaluation value by converting a result of the rating on the rating scale according to a predetermined conversion standard.

  Furthermore, by defining a conversion standard and deriving an evaluation value using this, it is possible to quantify the magnitude of the difference in voice quality on hearing.

  The voice quality difference evaluation table creation device according to the second aspect of the present invention is a device that creates a voice quality difference evaluation value table representing voice quality differences among a plurality of types of speech data. This voice quality difference evaluation table creation device includes an extraction means for extracting each of speech voice data included in a plurality of types of speech voice data in an order determined by a predetermined procedure, and a speech quality data extracted by the extraction means. Perceptual test on auditory voice quality is performed for each, and perceptual test means for deriving evaluation value on voice quality of the speech data, and evaluation on voice quality of multiple types of speech data obtained by the perceptual test A table creating means for creating a voice quality evaluation value table by storing an evaluation value related to a voice quality difference between arbitrary combinations of speech voice data based on the value in association with the combination of the speech voice data; including.

  In order to allow the subject to evaluate the audible impressions related to the audio stimulus, a more diverse evaluation can be obtained by a perceptual test. In addition, an evaluation value related to a voice quality difference can be derived with a small number of tests.

  Preferably, the perceptual test means includes a stimulus generation means for generating a voice stimulus based on each of the utterance voice data extracted by the extraction means, and a plurality of predetermined numbers for the voice stimulus generated by the stimulus generation means. Evaluation value acquisition means for acquiring the evaluation value of the subject regarding each of the rating scales.

  By allowing a subject to perform a rating using a plurality of rating scales, it is possible to derive an evaluation value related to a difference in voice quality based on a wider variety of audible impressions.

  More preferably, the evaluation value acquisition means is a discrete value representing a plurality of predetermined stages for the evaluation values of the subject regarding each of the plurality of rating scales for the audio stimulus generated by the stimulus generation means. Means for acquiring from the subject as a method, and vector generating means for generating a vector obtained by the means for acquiring and having a discrete evaluation value of the subject for each of a plurality of rating scales as elements. .

  By generating a vector based on the evaluation by the subject, the evaluation regarding the impression on the auditory sense by the subject can be quantified.

  More preferably, the table creation means, for an arbitrary combination of a plurality of utterance voice data, calculates a distance between vectors obtained by the vector generation means for the utterance voice data belonging to the combination. And a means for storing the distance calculated by the distance calculation means in the voice quality difference evaluation table in association with the corresponding combination of speech voice data.

  By using the distance between vectors obtained from the vector generation means as an evaluation value related to the difference in voice quality, the difference in voice quality can be quantified based on multifaceted evaluation by the subject.

  A voice corpus voice quality difference evaluation table creation system according to a third aspect of the present invention includes a voice corpus including a plurality of types of utterance voice data recorded under a plurality of different recording conditions, and a plurality of types of voice corpus included in the voice corpus. A voice quality difference evaluation table creation device according to the first aspect or the second aspect of the present invention, which uses speech voice data as an input.

  With this voice quality difference evaluation table creation system, it is possible to perform processing based on evaluation related to voice quality differences in various voice processing techniques performed using a voice corpus.

  A speech synthesis system according to a fourth aspect of the present invention comprises a speech corpus including a plurality of types of speech data, each including a plurality of speech data that can be separated into speech segments, and predetermined input information, A speech unit having a predetermined relationship with the input information is selected from a speech corpus, and a speech unit is synthesized by connecting a means for extracting and a series of speech units extracted by the means for extracting. And a voice quality created by the voice quality difference evaluation table creation device according to any one of the first to third aspects of the present invention using as input the plurality of types of speech data included in the speech corpus A speech synthesis system including a difference evaluation table. The means for extracting in this speech synthesis system selects speech segments having a predetermined relationship based on evaluation values of voice quality differences between a plurality of types of speech data stored in the voice quality difference evaluation table. .

  This speech synthesis system can prevent connection of speech segments having a large difference in voice quality when speech is synthesized based on a speech corpus including a plurality of types of speech speech data. Therefore, it is possible to reduce deterioration of the sound quality of the synthesized speech caused by connecting speech segments having a large difference in voice quality.

  Hereinafter, a speech unit connection type speech synthesis system according to an embodiment of the present invention will be described with reference to the drawings.

[First Embodiment]
FIG. 1 is a block diagram showing a functional configuration of a system according to an embodiment of the present invention. Referring to FIG. 1, this speech unit selection system 70 is a speech unit for selecting a speech unit by evaluation using an evaluation function 72 similar to that used in the conventional speech synthesis system shown in FIG. It includes a piece selection unit 44 and a voice quality difference evaluation table creation device 100 that creates a voice quality difference evaluation table 108 that stores information representing voice quality differences between sessions, which is referred to by the evaluation function 72.

  The voice quality difference evaluation table creation apparatus 100 includes a speech corpus 102 that stores speech data as selection candidates when the speech segment selection system 70 selects a series of speech segments 60 used for speech synthesis, and a speech corpus 102. Is used to derive evaluation values related to voice quality differences between sessions by performing a perceptual test on the subject 106 using voice data in the voice corpus 102, and to summarize the derived evaluation values related to voice quality differences. And a voice quality difference evaluation device 104 that creates a voice quality difference evaluation table 108. The voice quality difference evaluation table 108 is referred to at the time of evaluation by the evaluation function 72.

  In the speech corpus 102, speech data 110A, 110B,..., 110N composed of speech waveform signals for one session recorded and recorded by a single speaker are stored for a plurality of sessions. Identification numbers 112A, 112B,..., 112N for identifying each session are assigned to the audio data for one session.

  FIG. 2 shows the configuration of the voice quality difference evaluation apparatus 104 in the form of a block diagram. Referring to FIG. 2, voice quality difference evaluation apparatus 104 is connected to voice corpus 102 and is presented to subject 106 based on the identification number of the voice data of each session stored in voice corpus 102. The voice corpus 102 is connected to the test processing unit 122 that performs a process of determining a pair combination, the voice corpus 102, and the test processing unit 122, and presents a stimulus pair to be presented to the subject 106 according to a command from the test processing unit 122. And a stimulus extraction unit 120 that extracts the data from voice waveform data.

  The voice quality difference evaluation apparatus 104 further reproduces the voice waveform data extracted by the stimulus extraction unit 120 and presents the voice stimulus to the subject 106, and the voice stimulus presented to the subject 106. A rating acquisition unit 126 that acquires a rating of the subject 106, a combination of stimulus pairs that is connected to the test processing unit 122 and the rating acquisition unit 126 and is determined by the test processing unit 122, and a rating of the subject 106 that the rating acquisition unit 126 acquires. And a table creation unit 128 that creates a voice quality difference evaluation table 108 based on the above.

  The table creation unit 128 has a function of sending a completion signal to the test processing unit 122 when the processing for one stimulus pair is completed. Moreover, the test process part 122 has a function which starts the process of the next stimulus pair, if a completion signal is received.

  FIG. 3 shows an example of the configuration of the voice quality difference evaluation table 108. 3, voice quality difference evaluation table 108 includes entries 130A, 130B,... For each session. Each entry stores an identification number of the session and an evaluation value indicating a difference in voice quality between the voice recorded in the session and the voice recorded in each other session. These evaluation values are obtained by integrating evaluation results obtained from a plurality of subjects 106 for each combination of sessions.

  The system according to the present embodiment operates as follows. Note that when the voice quality difference evaluation apparatus 104 according to the present embodiment performs a perceptual test, the subject 106 is selected in advance by an appropriate method. In addition, it is assumed that the subject has been given sufficient and appropriate teaching in advance.

  Referring to FIG. 1, the voice naturally spoken by the speaker is stored in the voice corpus 102 as voice data 110A, 110B,. These audio data 110A, 110B,..., 110N are assigned identification numbers, respectively. In order to make the evaluation of the voice quality difference as accurate as possible, the same sentence is read at the beginning of each session, and this is used for the evaluation of the voice quality or the difference in voice quality.

  Referring to FIG. 2, when voice quality difference evaluation device 104 is activated, test processing unit 122 reads voice data identification numbers 112A, 112B,..., 112N stored in voice corpus 102 from voice corpus 102. . The test processing unit 122 gives the read identification number to the table creation unit 128. The table creation unit 128 prepares to create the voice quality difference evaluation table 108 based on the given identification number. That is, the voice quality difference evaluation table 108 shown in FIG. 3 that does not contain data is created on a storage device (not shown). When preparation for creating the voice quality difference evaluation table 108 is completed, the table creation unit 128 gives a completion signal to the test processing unit 122.

  In response to receiving the completion signal from the table creation unit 128, the test processing unit 122 is a target for comparing the voice quality differences among the identification numbers 112A, 112B,..., 112N acquired from the speech corpus 102. The identification number of the audio data of two sessions is selected and given to the stimulus extraction unit 120 and the table creation unit 128. At this time, there are various methods for selecting which two sessions. Any method may be used as long as all pairs of sessions to be compared can be extracted.

  When the stimulus extraction unit 120 is given the identification number of the voice data of the two sessions, the stimulus extraction unit 120 specifies each voice data corresponding to the given identification number from the voice data stored in the voice corpus 102. The voice waveform data for one utterance for voice quality evaluation is extracted from the voice data. The stimulus extraction unit 120 applies the extracted voice waveform data to the stimulus presentation unit 124 as a pair.

  The stimulus presentation unit 124 reproduces the given pair of speech waveform data in a form of comparison, and presents it to the subject 106. When the presentation of the stimulus is finished, the stimulus presentation unit 124 gives a signal indicating that the presentation is finished to the rating acquisition unit 126.

  The subject 106 is preliminarily shown with a rating scale for evaluating the auditory difference between the two stimuli constituting the stimulus pair. FIG. 4 shows an example of the rating scale. Referring to FIG. 4, the rating scale 160 is a scale constituted by five categories. A rating value represented by a numerical value corresponding to the category is assigned to each category in advance. The rating scale 160 may be given by a printed rating sheet. Moreover, you may display with respect to a test subject with a display apparatus.

  When a pair of stimuli is presented from the stimulus presentation unit 124, the subject 106 compares the stimuli. Subject 106 further determines which category of rating scale 160 is most appropriate to represent the difference between them and selects that category. In the present embodiment, the rating acquisition unit 126 includes a general input device (for example, a keyboard), and the subject 106 presses a key corresponding to the selected category.

  With reference to FIG. 2, when the subject 106 selects a category, the rating acquisition unit 126 gives a rating value assigned to the category to the table creation unit 128. When there are a plurality of subjects 106, the rating acquisition unit 126 specifies the rating values corresponding to the selections made by the plurality of subjects 106, and the table creation unit 128 uses the average value of the specified rating values as an integrated rating value. To give.

  The table creation unit 128 adds the rating value given by the rating acquisition unit 126 to an item corresponding to the combination of identification numbers given by the test processing unit 122 in the voice quality difference evaluation table 108 prepared in advance by the table creation unit 128. Store. The table creating unit 128 gives a completion signal to the test processing unit 122 in response to the completion of the storage of the rating value in the voice quality evaluation table 108.

  In response to the completion signal, the test processing unit 122 selects the identification number of the audio data of another two sessions from the identification numbers acquired from the audio corpus 102 by the selection method described above, and the stimulus extraction unit 120, and This is given to the table creation unit 128.

  The voice quality difference evaluation apparatus 104 repeats the above-described operation, and performs a perceptual test on the subject 106 for all combinations of sessions. The table creation unit 128 stores the rating value given from the rating acquisition unit 126 in the voice quality difference evaluation table 108. When the storage of the rating values for all the combinations is finished, the voice quality difference evaluation device 104 finishes a series of operations.

  The voice quality difference evaluation table 108 created in this way is referred to by the speech unit selection unit 44 when the value of the evaluation function 72 is calculated by the speech unit selection unit 44 shown in FIG. That is, referring to FIG. 1, the speech unit selection unit 44 uses the evaluation function 72 in addition to the observable physical quantity, the session to which the candidate speech unit belongs, and the speech function immediately before that. A set of identification numbers of two sessions with the session to which the speech unit belongs is given as a variable. The evaluation function 72 outputs an evaluation value based on the given physical quantity and the rating value stored in the item of the voice quality difference evaluation table 108 corresponding to the given set of identification numbers. The evaluation value reflects the evaluation value related to the difference in voice quality for each session.

  The speech segment selection unit 44 determines a series of speech segments based on this evaluation value. As a result, the voice synthesized by connecting a series of speech units by the speech unit connection unit 46 shown in FIG. 8 has less audible variation in voice quality between the connected speech units. For this reason, the synthesized voice can be heard naturally with reduced discontinuity.

[Second Embodiment]
In the system according to the first embodiment, the subject 106 compares the voice waveform data recorded in two different sessions, and evaluates the difference in voice quality in the voice waveform data. However, the present invention is not limited to such an embodiment.

  The voice quality difference evaluation apparatus according to the second embodiment creates a feature vector representing the voice quality of each session by perceptual inspection based on the voice data for each session, instead of comparing the voice data of two sessions. To do. The difference in voice quality between sessions is expressed by the distance between session vectors. In the perceptual test according to the present embodiment, an evaluation word set including a plurality of evaluation word pairs for evaluating voice quality is prepared in advance and given to a subject. Based on this evaluation word set, the test subject evaluates an audible impression regarding the voice quality of the presented stimulus.

  FIG. 5 is a block diagram showing the configuration of the voice quality difference evaluation apparatus 204 according to this embodiment. Referring to FIG. 5, the voice quality difference evaluation apparatus 204 includes a test processing unit 222 that performs a process of determining, in a predetermined order, a session of audio data from which audio stimuli to be extracted to be presented to the subject 106 are extracted. A stimulus extraction unit 220 that extracts voice data for specific voice quality evaluation from the voice data of the session determined by the test processing unit 222, and a stimulus presentation that presents the subject 106 with the voice stimulus extracted by the stimulus extraction unit 220 And a rating acquisition unit 226 for the subject 106 to acquire an auditory impression rating using the evaluation word set described above for the voice stimulus presented by the stimulus presentation unit 224.

  The voice quality difference evaluation device 204 is further connected to the test processing unit 222 and the rating acquisition unit 226, and based on the session identification number determined by the test processing unit 222 and the assessment of the subject acquired by the rating acquisition unit 226, Based on the voice quality vector creation unit 228 for creating a voice quality vector representing the voice quality of the voice recorded in each session, the voice quality vector table 230 for storing the voice quality vector, and the voice quality vector stored in the voice quality vector table 230, an arbitrary session A voice quality difference calculation unit 232 that calculates an evaluation value related to a voice quality difference between them, and a table creation unit 234 that creates a voice quality difference evaluation table 236 based on the evaluation value calculated by the voice quality difference calculation unit 232. The voice quality difference evaluation table 236 has the same configuration as the voice quality difference evaluation table 108 according to the first embodiment shown in FIG.

  FIG. 6 shows an example of the evaluation word set given to the subject in the perceptual test according to the present embodiment. Referring to FIG. 6, evaluation word set 240 includes evaluation word pairs composed of a pair of adjectives that are synonyms (“tensioned-no tension”, “cloudy-clear”, “bright-dark”, etc. ). For these evaluation word pairs, seven levels of categories (“very”, “pretty”,..., “Very”, “very”) are given from the evaluation word on the left side to the evaluation word on the right side. In these categories, as in the category of the evaluation scale according to the first embodiment, a rating value consisting of a numerical value is assigned to each evaluation word pair.

  FIG. 7 shows an example of the configuration of a voice quality vector table according to the present embodiment. Referring to FIG. 7, voice quality vector table 230 is a table constituted by entries 280A, 280B,... Corresponding to each session. Each entry of the voice quality vector table 230 includes a session identification number item 272 and a voice quality vector 274 indicating the voice quality of the voice data of the session. The voice quality vector 274 is a vector whose components are numerical values obtained by quantifying the subject's ratings for each evaluation word pair in the evaluation word set shown in FIG.

  The system according to the present embodiment operates as follows.

  Referring to FIG. 5, first, test processing unit 222 reads the identification numbers of all sessions from voice corpus 102 and provides them to table creation unit 234 and voice quality vector creation unit 228. Voice quality vector creating section 228 stores the given identification number in voice quality vector table 230 (see FIG. 7), and prepares voice quality vector table 230. When the preparation of the voice quality vector table 230 is completed, the voice quality vector creation unit 228 gives a completion signal to the test processing unit 222.

  In response to receiving the completion signal from the voice quality vector creation unit 228, the test processing unit 222 selects the identification number of the speech data of one session from the identification numbers read from the speech corpus 102, and the stimulus extraction unit 220 And provided to the voice quality vector creation unit 228.

  When this identification number is given, the stimulus extraction unit 220 uses the voice waveform data for one utterance for voice quality evaluation described above from the voice data of the session corresponding to the given identification number stored in the voice corpus 102. Is extracted and given to the stimulus presenting unit 224.

  The stimulus presentation unit 224 reproduces the audio waveform data, and when the reproduction is completed, the stimulus presentation unit 224 gives a signal indicating completion of reproduction to the rating acquisition unit 226.

  An evaluation word set shown in FIG. 6 is given to the subject 106 in advance. When the subject 106 receives a voice stimulus from the stimulus presentation unit 224, for each evaluation word pair in the given evaluation word set, the auditory impression received by the stimulus belongs to which category of the rating scale categories. Determine whether. The subject 106 selects the category determined to be the most appropriate from the rating scale corresponding to the evaluation word pair. The subject 106 performs this selection for all the evaluation word pairs.

Referring to FIG. 5, the rating acquisition unit 226 acquires a rating by the subject 106 in response to a signal indicating completion of reproduction from the stimulus presentation unit 224. The rating acquisition unit 226 specifies the category selected by the subject 106 for each evaluation word pair, and gives the corresponding rating value to the voice quality vector creation unit 228, respectively. If subjects 1 06 is plural, evaluation acquisition unit 226, like the first embodiment, to determine, respectively, respectively from the selection of a plurality of subjects 106, specified average value of a plurality of evaluation values Is provided to the voice quality vector creation unit 228 as an integrated rating value.

  The voice quality vector creating unit 228 creates a voice quality vector having these rating values as components based on the rating values corresponding to the evaluation word pairs. The voice quality vector creation unit 228 stores the created voice quality vector in the voice quality vector item of the entry of the identification number given from the test processing unit 222 in the voice quality vector table 230. When the storage of the voice quality vector is completed, the voice quality vector creation unit 228 gives a completion signal to the test processing unit 222. The test processing unit 222 determines the next session in response to the completion signal.

  By repeating the above operation, voice quality vectors for all sessions are stored in the voice quality vector table 230.

  When the creation and storage of the voice quality vector is completed, the voice quality difference calculation unit 232 calculates the evaluation value regarding the voice quality difference between sessions as follows. That is, the voice quality difference calculation unit 232 reads the identification numbers and voice quality vectors of any two sessions from the voice quality vector table 230. The voice quality difference calculation unit 232 calculates the distance between the read voice quality vectors in a multidimensional space around the rating scales of each evaluation word pair in the evaluation word set based on the two sets of read voice quality vectors. To do. The distance between vectors calculated here may be, for example, the Euclidean distance between vectors. The voice quality difference calculation unit 232 gives the distance between the vectors calculated in this way to the table creation unit 234 together with the identification numbers for the two sessions read out.

  The table creation unit 234 stores the distance between vectors given by the voice quality difference calculation unit 232 in the item corresponding to the identification number given by the voice quality difference calculation unit 232 in the voice quality difference evaluation table 236. The voice quality difference evaluation table 236 has the same configuration as the voice quality difference evaluation table 108 shown in FIG.

  The voice quality difference calculation unit 232 and the table creation unit 234 execute a series of these operations for all session combinations, and create a voice quality difference evaluation table 236.

  The voice quality difference evaluation table 236 created in this way is incorporated as one evaluation measure in the calculation of the evaluation function 72 shown in FIG. 1, as in the first embodiment.

  Thus, in the system including the voice quality difference evaluation apparatus according to the second embodiment, the subject does not directly evaluate the magnitude of the voice quality difference between the voice data for each session, but by the evaluation word set related to the voice quality. A voice quality difference evaluation table is created based on the expressed auditory impression of the subject. Therefore, it becomes possible to obtain more diverse evaluation results from the subject.

  As a result, in the speech synthesized by connecting the speech units selected by the speech unit selection unit 44 based on this evaluation value, the voice qualities of the connected speech units are similar. Less discontinuity occurs. Therefore, the synthesized voice is natural. Moreover, since it is sufficient to evaluate only once for each session, the time required for evaluation can be shortened as compared with the first embodiment in which evaluation is performed for each combination of two sessions.

  Each functional unit described in the above block diagram format can be realized by computer hardware and a program executed on the computer. As this computer, a computer having general-purpose hardware can be used as long as it has equipment for handling sound. Further, each functional block of the apparatus described above can be realized by a program by those skilled in the art based on the description in this specification. Such a program is also a piece of data and can be stored in a storage medium and distributed.

  Note that the number of subjects in the above-described embodiment does not matter. A test result with sufficient accuracy can be obtained by a subject who has been sufficiently taught and trained, even for a single subject or a small number of subjects. Further, by performing a perceptual test on a large number of subjects, it is possible to create a voice quality difference evaluation table based on a rating value close to that of a general user who hears synthesized speech.

  In the above-described embodiment, when a perceptual test is performed using a plurality of subjects, the rating values are integrated by the average value of the rating values obtained from each subject. However, the present invention is not limited to such an embodiment. In addition to the average value, the rating values may be integrated by the median value, the maximum value, the minimum value, and the like.

  In the above-described embodiment, the stimulus extraction units 120 and 220 extract the voice waveform data for one specific utterance from the voice data. However, the present invention is not limited to such an embodiment. For example, a plurality of utterances may be used. Although it is possible to evaluate using unspecified utterances, in that case, it is desirable to use a perceptual test method in which the accuracy of the evaluation is sufficiently guaranteed.

  In the above-described embodiment of the invention, a rating value is assigned in advance to each category of the rating scale given to the subject. However, the present invention is not limited to such an embodiment. For example, another perceptual test may be performed in advance to determine a rating value corresponding to each category. It is also possible to statistically determine the rating value of each category based on the test result obtained by the perceptual test.

  In the embodiment of the invention described above, as a perceptual test method and a rating value determination method for evaluating a voice quality difference, the subject is presented with a voice stimulus, and the voice stimulus is evaluated by a category scale. Met. However, the present invention is not limited to such an embodiment. Any method may be used as long as the extracted speech waveform data is used as a presentation stimulus and a scale based on the impression of hearing is formed by a perceptual test.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram which shows the structure of the system which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the voice quality difference evaluation apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows an example of the voice quality difference evaluation table 108 which concerns on embodiment of this invention. It is a figure which shows an example of the rating scale which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the voice quality difference evaluation apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows an example of the rating scale which concerns on the 2nd Embodiment of this invention. It is a figure which shows an example of the voice quality vector table which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the basic composition of a speech unit connection type | mold speech synthesis system.

Explanation of symbols

40,102 speech corpus, 42 input text, 44 speech unit selection unit, 46 speech unit connection unit, 48 synthesized speech data, 50, 72 evaluation function, 70 speech unit selection system, 100 voice quality difference evaluation table creation device, 104,204 Voice quality difference evaluation device, 106 subjects, 108, 236 Voice quality difference evaluation table, 110A, 110B,..., 110N Voice data, 112A, 112B,. Processing unit, 124, 224 Stimulus presentation unit, 126, 226 Rating acquisition unit, 128, 234 Table creation unit, 228 Voice quality vector creation unit, 230 Voice quality vector table, 232 Voice quality difference calculation unit

Claims (8)

A voice quality difference evaluation table creation device for creating a voice quality difference evaluation value table representing a voice quality difference between a plurality of types of utterance voice data, wherein each of the plurality of types of utterance voice data is used for evaluating a voice quality difference. Voice waveform data is included,
A combination determining means for determining an arbitrary combination of the first and second utterance voice data from the plurality of types of utterance voice data;
From each of said first and second speech data of the combination determined by the combination determining means, and the audio waveform data extracting means for extracting the speech waveform data for evaluation of differences in the voice,
The pre-Symbol Evaluation first and second audio stimulus consisting sound reproduced respectively audio waveform data for differences in voice quality of the first and second speech data, presented to the subject in comparison, the first and an input for assessing the magnitude of the difference in voice quality on the auditory sense of the second sound stimulus subjected to sensory test to receive from the subject, from the value of the rating entered, the voice quality during the speech data of the combination differences and a perception test means for deriving an evaluation value representing the evaluation of the size numeric,
The rating input indicates to which of a plurality of categories the magnitude of the voice quality difference represents the magnitude of the voice quality difference,
The evaluation value derived by the perceptual testing means, by storing in association with a combination of the speech data, and a table creation means for creating the voice quality difference evaluation value table creating voice difference evaluation table apparatus. The sensory test means comprises:
Voice waveform data for evaluating the difference in voice quality between the first and second utterance voice data is reproduced, and the first and second voice stimuli composed of the reproduced voices are subjected to a subject to be subjected to a perceptual test. and the stimulus presentation means for presenting in contrast Te,
The voice quality difference evaluation table creation according to claim 1, further comprising: an input means for receiving an input of a rating related to the magnitude of the difference in auditory voice quality between the first and second voice stimuli by the subject. apparatus. The input means includes
Means for causing the subject to input a rating indicating which of the plurality of categories the magnitude of the audible difference between the first and second audio stimuli belongs ;
Wherein Ri by the subject, the size of the audibility of difference between the first and second audio stimulus, response to the rating that indicates belongs to the plurality of stages of categories is input, the The voice quality difference evaluation table according to claim 2, further comprising means for outputting a numerical value assigned in advance to the category indicated by the input as an evaluation value of the difference between the first and second voice stimuli. Creation device. A voice quality difference evaluation table creation device for creating a voice quality difference evaluation value table representing a voice quality difference between a plurality of types of utterance voice data, wherein each of the plurality of types of utterance voice data is used for evaluating a voice quality difference. Voice waveform data is included,
An extracting means for extracting each of the voice waveform data for evaluation of the difference in voice quality included in the plurality of types of utterance voice data in an order determined by a predetermined procedure;
For each of the voice waveform data extracted by the extraction means, the voice quality on the auditory sense at the time of reproduction, performs perceptual tests to assess the subject for each of a plurality of different rating scale regarding voice quality, by the subject By deriving a plurality of evaluation values corresponding to the plurality of rating scales regarding the voice quality of the utterance voice data from the rating, a voice quality vector having the plurality of evaluation values as elements is obtained for each of the plurality of utterance voice data. A perceptual test means for calculating
As a distance defined between the voice quality vectors calculated for each of the plurality of types of utterance voice data by the perceptual test , evaluation regarding a voice quality difference between any combination of the plurality of types of utterance voice data A voice quality difference evaluation table creation device including table creation means for creating a voice quality difference evaluation value table by calculating a value and storing it in association with the combination of the speech data. The sensory test means comprises:
Reproduction means for presenting to the subject the voice obtained by reproducing each of the voice waveform data for evaluation of the difference in voice quality extracted by the extraction means;
Against the voice that has been presented to the subject by the reproducing means, viewed contains a commentary constant value acquisition means for acquiring a commentary constant value of the subject for each of the plurality of rating scale,
Each of the plurality of rating scales is represented by a plurality of categories,
The rating value of the test subject is, for each of the plurality of rating scales, which of the plurality of categories in which the voice quality of the voice presented to the subject by the playback means for the rating scale is the rating scale. Is a value indicating whether it belongs to
Each rating value is assigned a corresponding rating value,
The perceptual test means further includes a vector creating means for calculating the voice quality vector using an evaluation value assigned to a category corresponding to a rating value of a subject relating to each of the plurality of rating scales as an element. Item 5. The voice quality difference evaluation table creation device according to Item 4 . The table creation means includes
For any combination of said plurality of speech data, a distance calculation means for exiting calculate the Euclidean distance between the vector obtained by the vector generating means with respect to speech data belonging to the combination,
The Euclidean distance calculated by said distance calculation means includes a combination of the corresponding speech data associating, and means for storing the magnitude Satoshi the voice difference evaluation table of differences voice between the speech data The voice quality difference evaluation table creation device according to claim 5 . A voice corpus consisting of multiple types of speech data recorded under different recording conditions;
An input of said plurality of types of speech data included in the speech corpus, seen including a voice difference evaluation table generation apparatus according to any one of claims 1 to 6,
A speech corpus voice quality difference evaluation table creation system, wherein each of the plurality of types of utterance voice data includes voice waveform data for evaluation of a voice quality difference. A speech corpus comprising a plurality of types of speech data, each including a plurality of speech data that can be separated into speech segments;
Obtains input information predetermined by an evaluation function for selecting a speech unit for synthesizing speech from the input information, and means for selecting from pre-Symbol speech corpus the sequence of speech units, and extracts ,
Connect a series of speech unit means for the extraction is extracted, and means for synthesizing speech,
A speech synthesis system including the voice quality difference evaluation table created by the voice quality difference assessment table creation device according to any one of claims 1 to 7 with the plurality of types of speech voice data included in the voice corpus as input. Because
Each of the plurality of types of utterance voice data includes voice waveform data for evaluation of voice quality difference,
The means for extracting synthesizes speech from the input information using an evaluation value of a voice quality difference between the plurality of types of utterance voice data stored in the voice quality difference evaluation table as an input of the evaluation function. A speech synthesis system that selects speech segments to perform.

Images