JP2003271171A - Method, device and program for voice synthesis - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice synthesis method and a voice synthesis device capable of obtaining a synthesized voice having high voice quality with small memory capacity. <P>SOLUTION: This voice synthesis method comprises an attribute vector applying step for applying an attribute vector including a predetermined attribute factor and paralanguage information to each element piece of a voice corpus; a clustering step for clustering element pieces; a cluster representative value calculating step for calculating a cluster representative value; an explanation vector generating step for generating an explanation vector to be associated with the cluster representative value; a target attribute vector generating step for generating a target attribute vector of an element piece unit of a synthesized voice; an optimum approximate coefficient calculating step for calculating an optimum approximate coefficient that can optimally approximate the target attribute vector with the explanation vector in each element piece of the voice corpus; and a synthesized voice element piece generating step for generating the element piece of the synthesized voice on the basis of the optimum approximate coefficient, represents an attribute common to element pieces in the same cluster with the explanation vector and applies the attribute to voice synthesis. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing method, a voice synthesizing apparatus, and a voice synthesizing program.

[0002]

2. Description of the Related Art As a conventional speech synthesizing method and speech synthesizing apparatus, for example, Japanese Patent Laid-Open No. 2000-250.
There is one disclosed in Japanese Patent No. 570. This conventional speech synthesis method will be described with reference to FIG.

In FIG. 10, the pitch pattern database 11 stores pitch pattern data in units of accent phrases. This pitch pattern data has a pitch value for each frame. The prosody category is also described in each pitch pattern data. The text to be composed is given in units of accent phrases.

First, in step S31, pitch pattern data belonging to a prosody category that is the same as the prosody category of the text to be synthesized is pitch pattern database 1.
1 is searched for. Then, step S
In 31, when the pitch pattern data exists in the pitch pattern database 11, the process proceeds to step S33,
If it does not exist, the process proceeds to step S34. Then
In step S33, pitch pattern data is selected from a prosody category equal to the prosody category of the text to be synthesized. On the other hand, in step S34, from among the prosody categories included in the pitch pattern database 11,
The prosody category whose pitch pattern shape is considered to be the closest to the prosody category of the text to be synthesized is estimated. Then, in step S35, the above-mentioned step S33 is performed.
Similarly, pitch pattern data from the estimated prosody category is selected. Then, in step S36,
The difference vector between the selected prosody category and the text prosody category is applied to the selected pitch pattern data and transformed. Next, in step S37, the pitch pattern data is linearly expanded / contracted in the time axis direction in mora units, whereby the time length correction is performed according to the given time length. Next, in step S38, for the height of each pitch pattern, the midpoint between the heights of the start point and the end point of the pitch pattern is the average value of the pitch-tone component heights obtained from the pitch-tone pitch component speech-tone component determination algorithm. Is decided.

As described above, the voice pattern can be synthesized by acquiring the pitch pattern data suitable for the text to be synthesized from the pitch pattern database 11.

[0006]

However, in such a conventional speech synthesizing method, the classification of the segments in the speech corpus is performed based on the prosody category composed of many attributes rather than the segments themselves. Therefore, there is a problem that the memory capacity required for speech synthesis becomes very large. In addition, when the variance of data belonging to the prosody category is large, the representative values of the prosody category may not be able to approximate all the data, and there is a problem that the synthetic voice has an impression of variation due to the connection of the representative values.

The present invention has been made in order to solve such a problem, and can reduce the memory capacity required for voice synthesis, and also provide high-quality synthetic voice even when the variance of data belonging to the prosody category is large. A speech synthesis method, a speech synthesis apparatus, and a speech synthesis program capable of generating

[0008]

According to the speech synthesis method of the present invention, an attribute vector containing a predetermined attribute factor and paralinguistic information is given to each segment of a speech corpus including a prosodic segment and a speech waveform segment. An attribute vector assigning step to
A clustering step of clustering the pieces to which the attribute vector is given, a cluster representative value calculation step of calculating a cluster representative value of the pieces belonging to each cluster obtained by the clustering, and a clustering value obtained by the clustering The explanation vector generation step of generating an explanation vector based on the attribute vector of the segment belonging to each cluster, the target attribute vector generation step of generating a target attribute vector of each unit of synthesized speech, and the target attribute vector An optimal approximation coefficient calculation step of calculating an optimal approximation coefficient that is optimally approximated by an explanation vector for each speech corpus segment; and a synthetic speech segment generation step of generating a synthetic speech segment based on the optimal approximation coefficient. Is included. With this configuration, each segment of the prosodic segment and the speech waveform segment is clustered and then shared by the segment in the same cluster by the explanation vector generated based on the attribute vector of the segment in the same cluster. The attributes will be expressed and applied to speech synthesis.

In the speech synthesis method of the present invention, the attribute vector assigned in the attribute vector assigning step includes an expression of whether or not each attribute factor exists for each of the predetermined attribute factors. With this configuration, the attributes of each segment of the prosodic segment and the speech waveform segment can be simply expressed.

The speech synthesis method of the present invention includes a step of performing clustering based on auditory detection limits in the clustering step. With this configuration, variations in each of the prosodic segment and the speech waveform segment are taken into consideration.

In the speech synthesis method of the present invention, in the explanation vector generation step, each attribute factor of a vector obtained by adding the attribute vectors of the segment belonging to each cluster for each cluster is set for each attribute factor. The method includes a step of generating a vector having a new attribute factor that is obtained by dividing the total number of the pieces belonging to the cluster and setting the vector as the explanation vector of each cluster. With this configuration, each element of the prosody element and the speech waveform element has many attribute factors taken into consideration for each element.

In the speech synthesis method of the present invention, in the explanation vector generation step, each attribute factor of a vector obtained by adding the attribute vectors of the segment belonging to each cluster for each cluster is used as the attribute in the speech corpus. The step of generating a vector having a new attribute factor divided by the total number of data in which the factor has occurred and using it as the explanation vector of each cluster is included. With this configuration, each element of the prosody element and the speech waveform element has many attribute factors taken into consideration for each element.

The speech synthesis method of the present invention includes the step of, in the cluster representative value calculating step, using the centroid of the cluster as a representative segment of the cluster. With this configuration, the representative value of the cluster is represented by the center of gravity of the cluster.

The speech synthesis method of the present invention includes the step of, in the cluster representative value calculation step, using the mode of the cluster as a representative segment of the cluster. With this configuration, the representative value of the cluster is represented by the mode value of the cluster.

The speech synthesis method of the present invention comprises an attribute vector assigning step of assigning an attribute vector containing a predetermined attribute factor and paralinguistic information to each segment of a speech corpus containing a prosodic segment and a speech waveform segment. , A clustering step of clustering the pieces to which the attribute vector is added, a cluster representative value calculating step of calculating a cluster representative value of the pieces belonging to each cluster obtained by the clustering, and a clustering value obtained by the clustering The explanation vector generation step of generating an explanation vector based on the attribute vector of the segment belonging to each cluster, the explanation vector attribute factor comparison step of comparing the attribute factors of the explanation vectors with each other, and the segment unit of the synthesized speech. Target attribute vector generation step of generating the target attribute vector of An optimal approximation coefficient calculation step of calculating an optimal approximation coefficient that is optimally approximated by a vector for each speech corpus segment; and a synthetic speech segment generation step of generating a synthetic speech segment based on the optimal approximation coefficient. Including, the explanation vector attribute factor comparison step is regarded as the same when all the explanation vectors generated in the explanation vector step have attribute factors that can be regarded as the same according to a predetermined statistical significance level. The attribute factor is removed from the attribute factors of the explanation vector and the attribute vector. With this configuration, it is possible to eliminate the occurrence of clusters having the same representative value, select the optimum segment from the speech corpus, and apply it to speech synthesis.

The speech synthesis method of the present invention is the same when the plurality of explanation vectors generated in the explanation vector generation step in the explanation vector attribute factor comparison step can be regarded as the same according to a predetermined statistical significance level. It is characterized in that the cluster groups related to the explanation vector regarded as above are merged into one cluster. With this configuration, it is possible to eliminate the occurrence of clusters having the same representative value, select the optimum segment from the speech corpus, and apply it to speech synthesis.

The speech synthesis method of the present invention is the same when the plurality of explanation vectors generated in the explanation vector generation step in the explanation vector attribute factor comparison step can be regarded as the same according to a predetermined statistical significance level. The step of obtaining the number of the explanation vectors regarded as the above, the step of calculating the base 2 logarithm of the number, and the number of attribute factors corresponding to the integer of the logarithmic calculation result And a step of newly adding the attribute vector, the attribute vector including the added attribute factor is reassigned to the segment in the attribute vector assigning step. With this configuration, it is possible to eliminate the occurrence of clusters having the same representative value, select the optimum segment from the speech corpus, and apply it to speech synthesis.

In the speech synthesis method of the present invention, an inner product of the target attribute vector of the synthesized speech unit generated in the target attribute vector generation step and the explanation vector of the cluster of the speech corpus is calculated. Product calculation process,
A segment selection step of selecting a representative segment of a cluster having an explanation vector that is the maximum inner product among the calculated inner products. With this configuration, a segment corpus from which the attribute vector redundancy is removed is generated and applied to speech synthesis.

In the speech synthesis method of the present invention, the inner product of the target attribute vector of the synthesized speech unit generated in the target attribute vector generation step and the explanation vector of the cluster of the speech corpus is calculated. Product calculation process,
Calculating the sum of these inner products, and generating a synthesized speech unit by generating a weighted average of the representative units of the cluster with the value obtained by dividing each of the calculated inner products by the sum as a weight. The process is included. With this configuration, a segment corpus from which the attribute vector redundancy is removed is generated and applied to speech synthesis.

The speech synthesis method of the present invention is an optimum approximation for optimally approximating the target attribute vector of the synthesized speech unit generated in the target attribute vector generating step with the explanation vector of the cluster of the speech corpus. The method includes an optimum approximation coefficient calculation step of calculating a coefficient, and a synthetic speech element generation step of generating a synthetic speech element by weighted averaging representative speech elements based on the calculated optimum approximation coefficient. With this configuration, a segment corpus from which the attribute vector redundancy is removed is generated and applied to speech synthesis.

The speech synthesizer of the present invention comprises a representative segment storage means for storing a representative segment of a cluster of segments from a speech corpus including a prosodic segment and a speech waveform segment, and an explanation vector of the representative segment. An explanation vector storage means for storing
Pointer storage means for storing a pointer indicating the correspondence between the representative segment and the explanation vector, a text input means for inputting a text, a paralinguistic input means for inputting a paralanguage, and an analysis of the input text. Text analysis means, target attribute vector generation means for generating a target attribute vector for each unit of synthesized speech based on the analysis result of the text analysis means and the input paralinguistic information, and the generated target attribute An inner product calculating means for calculating an inner product of a vector and all the explanation vectors; an inner product maximum value segment selecting means for selecting a representative prosodic segment and a representative speech waveform segment which gives the maximum value of the inner product; The speech waveform segment transforming means for transforming the speech waveform segment according to the prosody segment thus modified and the transformed speech waveform segment are connected to each other. And a speech waveform segment connecting means that. With this configuration, each segment of the prosodic segment and the speech waveform segment is clustered and then shared by the segment in the same cluster by the explanation vector generated based on the attribute vector of the segment in the same cluster. The attributes will be expressed and applied to speech synthesis.

The speech synthesis apparatus of the present invention comprises a representative segment storage means for storing a representative segment of a cluster of segments from a speech corpus including a prosodic segment and a speech waveform segment, and an explanation vector of the representative segment. An explanation vector storage means for storing
Pointer storage means for storing a pointer indicating the correspondence between the representative segment and the explanation vector, a text input means for inputting a text, a paralinguistic input means for inputting a paralanguage, and an analysis of the input text. Text analysis means, target attribute vector generation means for generating a target attribute vector for each unit of synthesized speech based on the analysis result of the text analysis means and the input paralinguistic information, and the generated target attribute An inner product calculating means for calculating an inner product of a vector and all the explanation vectors, and a segment weighted averaging for performing weighted averaging of the representative prosodic segment and the representative speech waveform segment based on the calculated inner product. Means, and speech waveform segment transforming means for transforming the speech waveform segment that has been weighted and averaged according to the prosodic segment that has been weighted and averaged. And a speech waveform segments connecting means for connecting the voice waveform segments with each other, which is the deformation. With this configuration, each segment of the prosody segment and the speech waveform segment is
After the clustering, the attribute common to the segments in the same cluster is expressed by the explanation vector generated based on the attribute vector of the segments in the same cluster and applied to the speech synthesis.

The speech synthesizer of the present invention comprises a segment storage means for storing a segment from a speech corpus including a prosodic segment and a speech waveform segment, and an explanation vector storage means for storing an explanation vector of the segment. , Pointer storage means for storing a pointer indicating the correspondence between the segment and the explanation vector, a text input means for inputting a text, a paralinguistic input means for inputting a paralanguage, and an analysis of the input text. A text analysis unit, a target attribute vector generation unit for generating a target attribute vector for each unit of a synthesized speech segment based on the analysis result of the text analysis unit and the input paralinguistic information; and the synthesized speech segment. Optimal approximation coefficient calculating means for optimally approximating the target attribute vector of a unit with the description vector of the segment, A segment weighted averaging means for performing weighted averaging of the prosodic segment and the speech waveform segment based on an appropriate approximation coefficient, and the speech averaged according to the weighted averaged prosodic segment. The speech waveform segment transforming means for transforming the waveform segment is provided, and the speech waveform segment connecting means for connecting the modified speech waveform segments to each other. With this configuration, each segment of the prosodic segment and the speech waveform segment is clustered and then shared by the segment in the same cluster by the explanation vector generated based on the attribute vector of the segment in the same cluster. The attributes will be expressed and applied to speech synthesis.

The speech synthesis program of the present invention stores, in a computer, a segment storing step of storing a segment from a segment database, an explanation vector storing step of storing an explanation vector of the segment, the segment and the description. A pointer storing step of storing a pointer indicating a vector correspondence relationship, a text input step of inputting text, a paralinguistic input step of inputting a paralanguage, a text analysis step of analyzing the input text, and the text A target attribute vector generating step of generating a target attribute vector for each unit of synthesized speech based on the analysis result of the analysis step and the input paralinguistic information, the generated target attribute vector and all the explanation vectors And an inner product calculation step of calculating an inner product of and a prosodic segment and a speech waveform segment giving the maximum value of the inner product. Inner product maximum value segment selection step, a speech waveform segment transformation step of transforming the speech waveform segment in accordance with the selected prosodic segment, and a speech connecting the transformed speech waveform segments. It is for performing the waveform element connecting step. With this configuration, the computer clusters each segment of the prosodic segment and the speech waveform segment, and then the segment in the same cluster is identified by the explanation vector generated based on the attribute vector of the segment in the same cluster. Common attributes are expressed and applied to speech synthesis.

[0025]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below.

(Embodiment 1) A speech synthesis method according to a first embodiment of the present invention will be described with reference to the flowchart of FIG. First, in step S101, an attribute vector is assigned to each segment of the speech corpus. Here, the attribute vector is, for example, one automatically obtained from the language analysis and paralinguistic information. Examples of what is automatically obtained from the former linguistic analysis include part-of-speech, dependency, destination, mora number of accent phrase, and accent type, and examples of the latter paralinguistic information include tone, utterance style, There are things such as emotions that humans hear and judge. Table 1 shows an example of this attribute vector. In this attribute vector, 1 is entered when the element corresponds to the attribute factor, and 0 is entered when it does not correspond to the attribute factor, as in _ak shown in the following equation (1).

[Equation 1] However, k = 1, 2, 3, ..., N N: the number of data belonging to the cluster δ _ki = 1: when the attribute corresponds to the element specified by δ _ki = 0: does not correspond to the attribute specified by the element If

Next, in step S102, the speech corpus segment is clustered by the auditory detection limit. Here, the auditory detection limit means, for example, a time of about 5 msec in the case of a kind of phoneme duration of a prosodic unit, and an average spectrum difference of about 3 dB in the case of a voice unit. Next, in step S103, an explanation vector is given to each cluster. This description vector is defined as in Expression (2) and is associated with the representative value of the cluster.

[Equation 2] However,

[Equation 3] Further, r _i is the normalization coefficient of the i-th element and represents the total number of data in the cluster.

As shown in the above equation (2), the segment database is constructed by the explanation vector with the representative value of the cluster as the center of gravity.

Next, in step S104, language processing, speech style, tone instruction, etc. are input.
Next, in step S105, the target vector g _j of the synthesized speech in unit of the speech unit is generated by the language processing, the utterance style, the tone instruction, and the like input in S104 described above. Then, in step S106, equation (4)
As shown in, the inner product of the target attribute vector of the segment and the explanation vector is calculated.

[Equation 4] Then, in step S107, the above-mentioned step S
The inner product p ₁ calculated in 106 is normalized by the sum c of inner products to obtain the coefficient w ₁ .

[Equation 5] Here, the sum c of the inner products can be expressed by equation (6).

[Equation 6] Further, in step S107, to calculate the coefficients w ₁ described above in each segment c ₁ of the unit database, by adding by multiplying coefficients w ₁ to segment c _1, synthesized speech as shown in Equation (7) Generate a unit u _k of

[Equation 7] Next, in step S108, if the unit is a voice unit, the unit is transformed according to the target frequency or the phoneme duration, and if the unit is prosodic, the units are connected smoothly by a taper window to obtain a desired unit. Synthetic voice is generated.

As described above, according to the speech synthesis method of the present embodiment, each of the prosodic segment and the speech waveform segment is
After clustering, the attribute common to the segments in the same cluster is expressed by the explanation vector generated based on the attribute vector of the segments in the same cluster.
It is possible to generate a synthetic voice with a small memory capacity, and since it is possible to reduce the variation between units, it is possible to generate a high-quality synthetic voice.

By programming the above-described speech synthesis method, the computer can be made to execute the speech synthesis of this embodiment.

(Embodiment 2) A speech synthesis method according to a second embodiment of the present invention will be described with reference to the flowchart of FIG. First, in step S101, an attribute vector is assigned to each segment of the speech corpus. Here, the attribute vector is, for example, one automatically obtained from the language analysis and paralinguistic information. Examples of what is automatically obtained from the former linguistic analysis include part-of-speech, dependency, destination, mora number of accent phrase, and accent type, and examples of the latter paralinguistic information include tone, utterance style, There are things such as emotions that humans hear and judge. Table 1 shows an example of this attribute vector. In this attribute vector, 1 is entered when the segment corresponds to the attribute factor, and 0 is entered when the segment does not correspond to the attribute factor, as in the case of a _k shown in the above equation (1).

Next, in step S102, the speech corpus segment is clustered by the auditory detection limit. Here, the auditory detection limit means, for example, a time of about 5 msec in the case of a kind of phoneme duration of a prosodic unit, and an average spectrum difference of about 3 dB in the case of a voice unit. Next, in step S103, an explanation vector is given to each cluster. This description vector is defined as in the above equation (2) and is associated with the representative value of the cluster. As shown in the above equation (2), the segment database is constructed by the explanation vector with the representative value of the cluster as the center of gravity. Next, in step S104, language processing, speech style, tone instruction, etc. are input. Then
In step S105, by the language processing, the utterance style, the tone instruction, etc. input in S104 described above,
A target vector g _j of the synthesized speech in units of units is generated.
Next, in step S106, the inner product of the target attribute vector of the segment and the explanation vector is calculated as shown in the above equation (4).

Next, in step S201, of the inner products obtained in step S106 described above, the segment giving the maximum inner product is set as the segment for synthetic speech generation. That is,
Since the segment that maximizes the inner product has the attribute closest to the target attribute in the synthetic speech, it is the segment u _k used in the synthetic speech as shown in Expression (8). .

[Equation 8] Next, in step S108, if the unit is a voice unit, the unit is transformed according to the target frequency or the phoneme duration, and if the unit is prosodic, the units are connected smoothly by a taper window to obtain a desired unit. Synthetic voice is generated.

As described above, according to the speech synthesis method of the present embodiment, each of the prosodic segment and the speech waveform segment is
After the clustering, the attribute common to the segments in the same cluster is expressed by the explanation vector generated based on the attribute vector of the segments in the same cluster, and
The segment that maximizes the inner product is the segment of the synthetic speech, and the attributes of each segment of the prosodic segment and the speech waveform segment are easily expressed, so that the synthetic speech can be generated with a small memory capacity. Furthermore, since the variation between the segments can be reduced, it is possible to generate high-quality synthesized speech.

The computer can be made to execute the voice synthesis of this embodiment by programming the voice synthesis method described above.

(Embodiment 3) A speech synthesis method according to a third embodiment of the present invention will be described with reference to the flowchart of FIG. First, in step S101, an attribute vector is assigned to each segment of the speech corpus. Here, the attribute vector is, for example, one automatically obtained from the language analysis and paralinguistic information. Examples of what is automatically obtained from the former linguistic analysis include part-of-speech, dependency, destination, mora number of accent phrase, and accent type, and examples of the latter paralinguistic information include tone, utterance style, There are things such as emotions that humans hear and judge. Table 1 shows an example of this attribute vector. In this attribute vector, 1 is entered when the segment corresponds to the attribute factor, and 0 is entered when the segment does not correspond to the attribute factor, as in the case of a _k shown in the above equation (1).

Next, in step S102, the speech corpus segment is clustered by the auditory detection limit. Here, the auditory detection limit means, for example, a time of about 5 msec in the case of a kind of phoneme duration of a prosodic unit, and an average spectrum difference of about 3 dB in the case of a voice unit. Next, in step S103, an explanation vector is given to each cluster. This description vector is defined as in the above equation (2) and is associated with the representative value of the cluster. As shown in the above equation (2), the segment database is constructed by the explanation vector with the representative value of the cluster as the center of gravity. Next, in step S104, language processing, speech style, tone instruction, etc. are input. Then
In step S105, by the language processing, the utterance style, the tone instruction, etc. input in S104 described above,
A target vector g _j of the synthesized speech in units of units is generated.
Then, in step S301, the coefficient w ₁ to w _n for best fit in the sense of least squares the target vector described above in the description vector of segments by the equation (9) is calculated.

[Equation 9] Next, in step S302, the above-mentioned optimum approximation coefficient is multiplied by the speech units and added, thereby generating a synthesized speech speech unit u _k as shown in Expression (10).

[Equation 10] Next, in step S108, if the phoneme is a phonetic phoneme, it is transformed according to the target frequency and the phoneme duration,
In the case of prosodic variation, the connected parts are smoothly connected by a tapered window to generate a desired synthesized speech.

As described above, according to the speech synthesis method of the present embodiment, each of the prosodic segment and the speech waveform segment is:
After the clustering, the attribute common to the segments in the same cluster is expressed by the explanation vector generated based on the attribute vector of the segments in the same cluster, and
Since the speech synthesis speech segment is generated by the optimal approximation coefficient that optimally approximates the target vector with the segment description vector, the speech synthesis can be generated with a small memory capacity, and the variation between the speech segments can be reduced. As a result, it is possible to generate high-quality synthetic speech.

The computer can be made to execute the speech synthesis of the present embodiment by programming the above speech synthesis method.

(Embodiment 4) The speech synthesis method according to the fourth embodiment of the present invention uses the representative value of the cluster described in the above-mentioned first to third embodiments as the center of gravity. Then, the representative value of the cluster is used as the mode value of the cluster.

As an example of setting the representative value of the cluster as the mode value of the cluster, the data of the phoneme duration included in the prosodic segment will be described. Table 2 shows an example of typical values of phoneme duration, and the phoneme duration is shown on the left side of the table.
The right side of the table shows the number of data in a certain cluster.
For example, it is shown that the number of data in a cluster having a phoneme duration of 12 msec is two and the number of data in a cluster having a phoneme duration of 15 msec is one.
In Table 2, the center of gravity is 18.4 msec when the average value is calculated from the phoneme duration and the number of phonemes. In addition, since the number of pieces of data having a phoneme duration of 20 msec is the largest of 10, the mode value is 20 msec. When exceptional data is present in the data in the cluster as shown in Table 2, a more stable speech synthesis can be generated by using the mode value rather than the centroid point. Instead of using the representative value of the cluster described in the first to third embodiments as the center of gravity, it is preferable to use the representative value of the cluster as the mode value of the cluster. On the other hand, even when there is no biased data in the cluster, the mode value and the center of gravity often coincide with each other. Therefore, it is generally preferable to use the mode value as the representative value of the cluster.

As described above, according to the present embodiment, the representative value of the cluster is represented by the mode value of the cluster.
It is possible to generate stable synthetic speech.

By programming the above-mentioned speech synthesis method, the computer can be made to execute the speech synthesis of this embodiment.

(Embodiment 5) A speech synthesis method according to a fifth embodiment of the present invention will be described with reference to the flowchart of FIG. The speech synthesis method according to the present embodiment relates to the processing when generating the attribute vector and the explanation vector described in the first to third embodiments. Instead of the normalization factor r ₁ ~r _n elements of description vectors of the aforementioned segment database, it is an occurrence data the total number of the factors in the original speech corpus.

First, in step S401, the i-th element of the explanation vectors is compared. Next, in step S402, it is determined by a predetermined statistical method, for example, a chi-square test, whether the difference between the i-th elements of the explanation vectors compared in step S401 is significantly large. Here, the criterion of judgment is based on a predetermined threshold value. When the difference of the i-th element is significantly large, the process proceeds to step S404, and the element to be compared is advanced. On the other hand, when the difference in the i-th element is not significantly large, the process proceeds to step S403, and it is determined whether all the explanation vectors have been compared. When all the explanation vectors have been compared, the process proceeds to step S406, and when all the explanation vectors have not been compared, the process proceeds to step S405 and the compared explanation vector is replaced.

Next, in step S406, the i-th element whose difference is not significantly large is excluded from the attribute vector and the explanation vector. In other words, if they can be regarded as the same at a predetermined significance level,
The attribute of this element means that it occurs at a common frequency in each cluster, and does not contribute to the formation of clusters. Therefore, this element is not applied to the explanation vector and the attribute vector as a meaningless element. Next, in step S407, it is determined whether or not the comparison of all the elements is completed. If the comparison of all the elements has not been completed, the process proceeds to step S408, and the element to be compared is advanced to the next, and if the comparison of all the elements is completed, the process ends.

As described above, according to the present embodiment, the elements which can be regarded as the same at the predetermined statistical significance level are not applied as the meaningless elements to the explanation vector and the attribute vector, and the explanation is made. Since the vector and the attribute vector can be set to the optimum size, it is possible to generate a synthesized voice with a small memory capacity, and further, since the variation between the units can be reduced, it is possible to generate a high-quality synthesized voice.

By programming the above-described speech synthesis method, it is possible to cause a computer to execute the speech synthesis of this embodiment.

(Sixth Embodiment) A speech synthesis method according to a sixth embodiment of the present invention will be described with reference to the flowchart of FIG. The speech synthesis method according to the present embodiment relates to the handling of the explanation vector described in the first to third embodiments. Instead of the normalization factor r ₁ ~r _n elements of description vectors of the aforementioned segment database, it is an occurrence data the total number of the factors in the original speech corpus.

First, in step S501, the explanation vectors are compared with each other. Next, in step S502, it is determined whether or not the difference between the explanation vectors compared in step S501 is significantly large by a predetermined statistical method, for example, the chi-square test. Here, the criterion of judgment is based on a predetermined threshold value. If the difference between the explanation vectors is significantly large, step S50.
4, the explanation vector to be compared is advanced next. On the other hand, when the difference between the explanation vectors is not significantly large, the process proceeds to step S503, and it is determined whether all the explanation vectors have been compared. If all the explanation vectors have been compared, the process proceeds to step S506, and if all the explanation vectors have not been compared, step S505.
The explanation vector to be compared is replaced. Then
In step S506, clusters in which the difference between the explanation vectors is not significantly large are integrated and a new segment database is constructed. In other words, if they can be regarded as the same at a predetermined significance level,
This means that the clustering by the auditory detection limit shown in the first embodiment of the present invention described above is too strict, and the cluster is forcibly divided because the variation of the data of the original segment is large. It is conceivable that the size of the phoneme database can be reduced by aggregating the phoneme data that can be regarded as the same at a predetermined constant significance level and using the average value as the representative value of the cluster.

As described above, according to the present embodiment, the size of the segment database can be reduced by integrating the clusters of explanation vectors that can be regarded as the same at a predetermined statistical significance level. The synthetic speech can be generated with a small memory capacity, and further, the variation between the segments can be reduced, so that the synthetic speech with high sound quality can be generated.

It should be noted that, due to the above-mentioned processing relating to the handling of the explanation vector, there may be a case where the variation between the representative values exceeds the auditory detection limit, which may result in a synthetic speech having a rough sound quality. Whether or not to reduce the number can be determined while checking the viewing.

The computer can be made to execute the voice synthesis of this embodiment by programming the above-mentioned voice synthesis method.

(Embodiment 7) A speech synthesis method according to a seventh embodiment of the present invention will be described with reference to the flowchart of FIG. The speech synthesis method according to this embodiment relates to the attribute factors described in the above-described first to third embodiments. Instead of the normalization factor r ₁ ~r _n elements of description vectors of the aforementioned segment database, it is an occurrence data the total number of the factors in the original speech corpus.

First, in step S601, two explanation vectors are arbitrarily selected. Then, step S
At 602, the i th element is compared between the description vectors. Next, in step S603, i between the explanation vectors compared in step S602 is determined by a predetermined statistical method, for example, a chi-square test.
It is determined whether the difference of the th element is significantly large.
Here, the criterion of judgment is based on a predetermined threshold value. If the difference in the i-th element between the explanation vectors is significantly large, the process proceeds to step S604, and if the difference in the i-th element between the explanation vectors is not significantly large, the process proceeds to step S609. Next, in step S604, it is determined whether the comparison with all the explanation vectors has been completed. If the comparison with all the explanation vectors is completed, step S606.
If the comparison with all the explanation vectors is not completed, the process proceeds to step S605. Then, step S60
In step 6, the integer number of Log ₂ N becomes the number of new additional attributes based on the number N of the explanation vectors that the difference of the i-th element between the explanation vectors is significantly large in step S603. To be done. Next, in step S607, new additional attributes are determined by comparative observation of the data belonging to the N clusters described above, the attribute vectors of these data are reassigned and updated, and as a result, the explanation vector is also updated. It is like this. Next, in step S608, the presence / absence of the above-described attribute is reviewed in all data of other clusters with respect to the new additional attribute, the attribute vector is reassigned and updated, and as a result, the explanation vector is also updated. Has become. When the process of step S608 ends, step S602
To return to.

On the other hand, in step S603 described above,
If it is determined that the difference in the i-th element between the description vectors is not significantly large, the element to be compared is advanced in step S609, and the process proceeds to step S610. Next, in step S610, it is determined whether the comparison has been completed for all the elements. If the comparison has not been completed for all elements, the process returns to step S602, and if the comparison has been completed for all elements, the process ends. In addition, the above step S
In 604, when the comparison with all the explanation vectors has not been completed, the process proceeds to step S605, the explanation vector to be compared is replaced, i = 1, and the process returns to step S602.

As described above, according to the present embodiment, it is possible to steadily search for the attribute factor that has been overlooked by comparing the i-th elements between the explanation vectors, and construct a more effective database. Therefore, it is possible to generate a voice synthesis with good sound quality.

The computer can be made to execute the speech synthesis of this embodiment by programming the speech synthesis method described above.

(Embodiment 8) A speech synthesizer according to an eighth embodiment of the present invention will be described with reference to the block diagram of FIG. First, the configuration of the speech synthesizer of this embodiment will be described. As shown in FIG. 7, the speech synthesizer according to the present embodiment has text input means 701 for inputting text.
A text analysis means 703 for analyzing the text, a target vector generation means 704 for generating a target vector,
Speech waveform segment transforming means 705 for transforming speech waveform segments.
And a voice waveform segment connecting means 7 for connecting the voice waveform segment
06 and para-language input means 702 for inputting a para-language
And the explanation vector storage means 70 for storing the explanation vector.
7 and the inner product calculating means 708 for calculating the inner product of the vector
A pointer storage unit 709 for storing a pointer, an inner product maximum value segment selection unit 710 for selecting a segment to which the maximum inner product is given, and a segment storage unit 711 for storing the segment.

Next, the operation of the speech synthesizer of this embodiment will be described. First, the text is text input means 701.
Entered in. Next, after being inputted to the text analysis means 703, it is inputted to the target vector generation means 704.
On the other hand, the paralinguistic information is input to the paralinguistic input unit 702 for each unit such as a phrase or an accent phrase, and then to the target vector generating unit 704. This paralinguistic information may be determined by a human or may be uniquely determined from a sentence pattern. Next, the target vector generation means 704
Then, an attribute vector expressing the attribute of each voice unit to be synthesized is generated.

On the other hand, the segment stored in the segment storage unit 711 and the explanation vector stored in the explanation vector storage unit 707 are associated with each other by the pointer stored in the pointer storage unit 709. Further, the inner product of the attribute vector and all the explanation vectors of the unit database is calculated for each unit by the inner product calculating unit 708, and the inner product maximum value unit selecting unit 710 refers to the pointer of the unit giving the maximum inner product. Then, the segment giving the maximum inner product is selected from the prosodic segment database and the speech waveform segment database. Of the selected speech segments, the speech waveform segment changing means 70 matches the speech waveform segment according to the prosody segment.
Transformed in 5. Next, the speech waveform segments are connected to each other by the speech waveform segment connection means, and the synthetic speech is generated.

As described above, according to the speech synthesizer of this embodiment, the segment that maximizes the inner product of the target vector of the input text and the explanation vector generated based on the attribute vector of the segment. Is calculated and used as a unit of the synthesized voice, a unit closest to the target attribute of the synthesized voice can be obtained, and a synthesized voice of high sound quality can be generated.

(Ninth Embodiment) A speech synthesizer according to a ninth embodiment of the present invention will be described with reference to the block diagram of FIG. First, the configuration of the speech synthesizer of this embodiment will be described. As shown in FIG. 8, the speech synthesizer according to the present embodiment has a text input means 701 for inputting text.
A text analysis means 703 for analyzing the text, a target vector generation means 704 for generating a target vector,
Speech waveform segment transforming means 705 for transforming speech waveform segments.
And a voice waveform segment connecting means 7 for connecting the voice waveform segment
06 and para-language input means 702 for inputting a para-language
And the explanation vector storage means 70 for storing the explanation vector.
7 and the inner product calculating means 708 for calculating the inner product of the vector
A pointer storage means 709 for storing pointers, a segment weighted averaging means 801 for averaging by a weighting coefficient normalized by dividing by the sum of inner products, and a segment storage means 711 for storing segments. ing.

Next, the operation of the speech synthesizer of this embodiment will be described. First, the text is text input means 701.
Entered in. Next, after being inputted to the text analysis means 703, it is inputted to the target vector generation means 704.
On the other hand, the paralinguistic information is input to the paralinguistic input unit 702 for each unit such as a phrase or an accent phrase, and then to the target vector generating unit 704. This paralinguistic information may be determined by a human or may be uniquely determined from a sentence pattern. Next, the target vector generation means 704
Then, an attribute vector expressing the attribute of each voice unit to be synthesized is generated.

On the other hand, the segment stored in the segment storage unit 711 and the explanation vector stored in the explanation vector storage unit 707 are associated with each other by the pointer stored in the pointer storage unit 709. Further, the inner product of the attribute vector and all the explanation vectors of the unit database is calculated for each unit by the inner product calculating unit 708, and is divided by the sum of the inner products by the unit weighted averaging unit 801 to obtain a normalized weighting coefficient. Averaged and weighted averaged pieces are generated. Next, the speech waveform segment transforming means 70
5, the speech waveform segments are connected to each other by the speech waveform segment connection means 706, and a synthetic speech is generated.

As described above, according to the speech synthesizer of the present embodiment, the inner product of the target vector of the input text and the explanation vector generated based on the attribute vector of the segment is calculated to obtain the segment. Since the weighted averaging is performed, it is possible to prevent the occurrence of clusters having the same representative value, and it is possible to reduce the variation between the segments, so that it is possible to generate high-quality synthetic speech.

(Embodiment 10) A speech synthesizer according to a tenth embodiment of the present invention will be described with reference to the block diagram of FIG. First, the configuration of the speech synthesizer of this embodiment will be described. As shown in FIG. 9, the speech synthesizer according to the present embodiment has a text input means 70 for inputting text.
1 and a text analysis means 703 for analyzing the text,
Target vector generation means 704 for generating a target vector
And a speech waveform segment transforming means 7 for transforming a segment of a speech waveform.
05, a voice waveform segment connecting means 706 for connecting a voice waveform segment, and a paralinguistic input means 70 for inputting a paralanguage.
2 and the explanation vector storage means 7 for storing the explanation vector
07, and an optimum approximation coefficient calculating means 9 for calculating an optimum approximation coefficient for optimally approximating the target vector with the segment description vector.
01 and pointer storage means 709 for storing a pointer
, A piece weighted averaging means 902 for weighted averaging the pieces by the optimum approximation coefficient, and a piece storage means 71 for storing the pieces.
1 and.

Next, the operation of the speech synthesizer of this embodiment will be described. First, the text is text input means 701.
Entered in. Next, after being inputted to the text analysis means 703, it is inputted to the target vector generation means 704.
On the other hand, the paralinguistic information is input to the paralinguistic input unit 702 for each unit such as a phrase or an accent phrase, and then to the target vector generating unit 704. This paralinguistic information may be determined by a human or may be uniquely determined from a sentence pattern. Next, the target vector generation means 704
Then, an attribute vector expressing the attribute of each voice unit to be synthesized is generated.

On the other hand, the segment stored in the segment storage unit 711 and the explanation vector stored in the explanation vector storage unit 707 are associated with each other by the pointer stored in the pointer storage unit 709. Further, the optimum approximation coefficient calculating unit 901 calculates an optimum approximation coefficient that optimally approximates the target vector with the explanation vector of the unit, and outputs it to the unit weighted averaging unit 902. Element weighted averaging means 902
In, a weighted averaged segment is generated by the optimum approximation coefficient. Next, the speech waveform segment is transformed by the speech waveform segment transforming unit 705, the speech waveform segments are connected to each other by the speech waveform segment connecting unit 706, and a synthetic speech is generated.

As described above, according to the speech synthesizer of the present embodiment, the optimum approximation coefficient for optimally approximating the target vector of the input text with the explanation vector generated based on the attribute vector of the segment is obtained. By calculating and performing weighted averaging of the speech units by this optimum approximation coefficient to obtain the speech synthesis speech units, it is possible to prevent the occurrence of clusters having the same representative value and reduce the variation between the speech units. Therefore, it is possible to generate high-quality synthetic speech.

[0072]

As described above, according to the present invention,
After the prosodic segment and the speech waveform segment are clustered, the attribute common to the segment in the same cluster is expressed by the explanation vector generated based on the attribute vector of the segment in the same cluster. As a result, it is possible to generate synthetic speech with a small memory capacity, and since it is possible to reduce the variation between units, it is possible to generate high-quality synthetic speech.

[Brief description of drawings]

FIG. 1 is a flowchart of a speech synthesis method according to a first embodiment of the present invention.

FIG. 2 is a flowchart of a speech synthesis method according to the second embodiment of the present invention.

FIG. 3 is a flowchart of a voice synthesis method according to a third embodiment of the present invention.

FIG. 4 is a flowchart of a speech synthesis method according to the fifth embodiment of the present invention.

FIG. 5 is a flowchart of a speech synthesis method according to the sixth embodiment of the present invention.

FIG. 6 is a flowchart of a voice synthesis method according to a seventh embodiment of the present invention.

FIG. 7 is a block diagram of a speech synthesizer according to an eighth embodiment of the present invention.

FIG. 8 is a block diagram of a speech synthesizer according to a ninth embodiment of the present invention.

FIG. 9 is a block diagram of a speech synthesizer according to a tenth embodiment of the present invention.

FIG. 10 is a flowchart of a conventional speech synthesis method.

[Explanation of symbols]

701 Text input means 702 Para-language input means 703 Text analysis means 704 target vector generation means 705 Speech waveform segment transforming means 706 Speech waveform segment connecting means 707 means for storing explanation vector 708 Inner product calculating means 709 pointer storage means 710 Inner product maximum value unit selection means 711 Element storage means 801, 902 Element weighted averaging means 901 Optimal approximation coefficient calculation means

Claims (17)

[Claims] 1. An attribute vector assigning step of assigning an attribute vector including a predetermined attribute factor and paralinguistic information to each segment of a speech corpus including a prosody segment and a speech waveform segment, and assigning the attribute vector. A clustering step of clustering the segment, the cluster representative value calculating step of calculating a cluster representative value of the segment belonging to each cluster obtained by the clustering, and the cluster belonging value obtained by the clustering An explanation vector generation step of generating an explanation vector based on the attribute vector of a segment, a target attribute vector generation step of generating a target attribute vector of a unit of synthetic speech in a unit of a speech unit, and the target attribute vector optimized by the explanation vector. An optimal approximation coefficient calculation step of calculating an approximate approximation coefficient for each segment of the speech corpus, A speech synthesis unit for generating synthesized speech units based on the optimum approximation coefficient. 2. The attribute vector assigned in the attribute vector assigning step includes an expression of whether or not each attribute factor exists for each of the predetermined attribute factors. The described speech synthesis method. 3. The speech synthesis method according to claim 1, wherein the clustering step includes a step of performing clustering based on an auditory detection limit. 4. The explanatory vector generating step includes adding each attribute factor of a vector obtained by adding attribute vectors of the segment belonging to each cluster for each cluster to the segment belonging to each cluster for each attribute factor. 2. The speech synthesis method according to claim 1, further comprising a step of generating a vector having a new attribute factor divided by the total number of s, and setting it as an explanation vector of each cluster. 5. The total number of data in which the attribute factors are generated in the speech corpus, wherein each of the attribute factors of the vector obtained by adding the attribute vectors of the segment belonging to each cluster is added to the explanation vector generation step. 2. The speech synthesis method according to claim 1, further comprising the step of generating a vector whose division factor is a new attribute factor and using it as an explanation vector of each cluster. 6. The speech synthesizing method according to claim 1, wherein the cluster representative value calculating step includes a step of using a center of gravity of the cluster as a representative segment of the cluster. 7. The speech synthesis method according to claim 1, wherein the cluster representative value calculating step includes a step of setting a mode value of the cluster as a representative segment of the cluster. 8. An attribute vector assigning step of assigning an attribute vector including a predetermined attribute factor and paralinguistic information to each segment of a speech corpus including a prosody segment and a speech waveform segment, and assigning the attribute vector. A clustering step of clustering the segment, the cluster representative value calculating step of calculating a cluster representative value of the segment belonging to each cluster obtained by the clustering, and the cluster belonging value obtained by the clustering An explanation vector generation step of generating an explanation vector based on the attribute vector of a segment, an explanation vector attribute factor comparison step of comparing each attribute factor of the explanation vectors with each other, and a target attribute vector of each synthesized speech segment unit. And a target attribute vector generation step for optimally approximating the target attribute vector with the explanation vector. An optimum approximation coefficient calculation step of calculating an optimum approximation coefficient for each speech corpus segment, and a synthetic speech segment generation step of generating a synthetic speech segment based on the optimal approximation coefficient. In the attribute factor comparison step, if there is an attribute factor that can be regarded as the same according to a predetermined statistical significance level common to all the explanation vectors generated in the explanation vector step, the attribute factor considered as the same is A method for synthesizing speech, which is characterized by excluding an explanation vector and an attribute factor of the attribute vector. 9. The explanation vector attribute factor comparison step,
When the plurality of explanation vectors generated in the explanation vector generation step can be regarded as the same according to a predetermined statistical significance level, clusters related to the same regarded explanation vectors are merged into one cluster. 9. The voice synthesis method according to claim 8, wherein: 10. The explanation vector attribute factor comparing step is regarded as the same when the plurality of explanation vectors generated in the explanation vector generating step can be regarded as the same according to a predetermined statistical significance level. A procedure for obtaining the number of vectors, a procedure for calculating a base 2 logarithm of the number, and a new number of attribute factors corresponding to the number obtained by converting the logarithmic calculation result into an integer are provisionally newly added to the segment. 9. The voice synthesizing method according to claim 8, further comprising the step of re-assigning an attribute vector including the added attribute factor to the segment in the attribute vector assigning step. 11. An inner product calculating step of calculating an inner product of the target attribute vector of the synthesized speech segment unit generated in the target attribute vector generating step and the explanation vector of the cluster of the speech corpus, 9. The speech synthesis according to claim 1, further comprising: a segment selection step of selecting a representative segment of a cluster having an explanation vector that is the maximum inner product among the calculated inner products. Method. 12. An inner product calculating step of calculating an inner product of the target attribute vector of the synthesized speech segment unit generated in the target attribute vector generating step and the explanation vector of the cluster of the speech corpus, And a step of generating a synthetic speech segment by performing a weighted average of the representative segments of the cluster with a value obtained by dividing each of the calculated inner products by the total as a weight. 9. The voice synthesis method according to claim 1, further comprising: 13. Optimal approximation for calculating an optimal approximation coefficient that optimally approximates the target attribute vector of the synthesized speech segment unit generated in the target attribute vector generation step with the explanation vector of the cluster of the speech corpus. 2. The method according to claim 1, further comprising: a coefficient calculating step; and a synthetic speech element generating step of generating a synthetic speech element by weighted averaging representative speech elements based on the calculated optimum approximation coefficient.
And the speech synthesis method according to claim 8. 14. A representative segment storage means for storing a representative segment of a cluster of segments from a speech corpus including a prosody segment and a speech waveform segment, and an explanation vector storage for storing an explanation vector of the representative segment. Means, pointer storage means for storing a pointer indicating the correspondence between the representative segment and the explanation vector, text input means for inputting text, para-language input means for inputting para-language, and the input text And a target attribute vector generation unit that generates a target attribute vector for each unit of synthesized speech based on the analysis result of the text analysis unit and the input paralinguistic information. An inner product calculating means for calculating an inner product of the target attribute vector and all the explanation vectors, and a representative prosodic segment giving the maximum value of the inner product. And an inner product maximum value segment selection means for selecting a representative speech waveform segment, a speech waveform segment transformation means for transforming the speech waveform segment according to the selected prosodic segment, and the transformed speech. A voice synthesizing apparatus comprising: a voice waveform segment connecting means for connecting waveform segments. 15. A representative segment storage means for storing a representative segment of a cluster of segments from a speech corpus including a prosodic segment and a speech waveform segment, and an explanation vector storage for storing an explanation vector of the representative segment. Means, pointer storage means for storing a pointer indicating the correspondence between the representative segment and the explanation vector, text input means for inputting text, paralanguage input means for inputting paralanguage, and the input text And a target attribute vector generation unit that generates a target attribute vector for each unit of synthesized speech based on the analysis result of the text analysis unit and the input paralinguistic information. An inner product calculating means for calculating an inner product of the target attribute vector and all the explanation vectors, and the representative based on the calculated inner product. A segment weighted averaging means for performing weighted averaging of the prosodic segment and the representative speech waveform segment, and transforming the weighted averaged speech waveform segment according to the weighted averaged prosody segment. A voice synthesizing apparatus comprising: a voice waveform segment deforming means; and a voice waveform segment connecting means for connecting the deformed voice waveform segments. 16. A segment storage unit for storing a segment, an explanation vector storage unit for storing an explanation vector of the segment, and a pointer storage unit for storing a pointer indicating a correspondence between the segment and the explanation vector. A text inputting means for inputting text, a paralinguistic inputting means for inputting paralinguistics, a text analyzing means for analyzing the inputted text, an analysis result of the text analyzing means and the inputted paralinguistic information. A target attribute vector generating means for generating a target attribute vector for each unit of synthesized speech units based on the above, and an optimum approximation for optimally approximating the target attribute vector of each unit of the synthesized speech by the explanation vector of the unit. Optimum approximation coefficient calculating means for calculating a coefficient, and weighted averaging of the prosodic element and the speech waveform element based on the optimum approximation coefficient. A speech segment weighted averaging means, and a speech waveform segment transformation means for transforming the speech waveform segment weighted and averaged according to the weighted averaged prosodic segment.
A voice synthesizing apparatus comprising: a voice waveform segment connecting means for connecting the deformed voice waveform segments. 17. A computer storing a segment storing step of storing a segment from a segment database, an explanation vector storing step of storing an explanation vector of the segment, and a correspondence relationship between the segment and the explanation vector in a computer. A pointer storing step of storing a pointer, a text inputting step of inputting text, a paralinguistic inputting step of inputting paralinguistics, and a text analyzing step of analyzing the input text,
A target attribute vector generation step of generating a target attribute vector for each unit of synthesized speech based on the analysis result of the text analysis step and the input paralinguistic information; and the generated target attribute vector and all of the above An inner product calculating step of calculating an inner product with the explanation vector, an inner product maximum value segment selecting step of selecting a prosodic segment and a speech waveform segment giving the maximum value of the inner product, and the selected prosodic segment A speech synthesizing program for executing a speech waveform segment transformation step of transforming the speech waveform segment accordingly and a speech waveform segment connection step of connecting the transformed speech waveform segments.