JP4403996B2 - Prosody pattern generation apparatus, prosody pattern generation method, and prosody pattern generation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To dynamically generate prosodic patterns with high naturalness to inputted pronunciation symbol strings. <P>SOLUTION: An attribute information extraction means 12 extracts attribute information of the inputted pronunciation symbol strings. A similarity calculation means 13 calculates distance between attribute information of the prosodic patterns in a prosodic data base 16 and attribute information of the pronunciation symbol strings to calculate weight to each prosodic pattern. A prosodic pattern generation means 14 generates a new prosodic pattern by continuously coupling the prosodic patterns by the calculated weight. A waveform generation means 15 controls prosody by the generated new prosodic pattern to generate waveform of a synthetic sound. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to a voice processing technique, and more particularly to a technique for generating a prosodic pattern.

  In text-to-speech synthesis, synthesized speech is generated by controlling the speech waveform with a prosodic pattern estimated for the input text. A prosodic pattern is a pattern of temporal changes such as voice pitch (fundamental frequency), length, and strength. Prosodic patterns have a great influence on the naturalness of synthesized sounds. In speech synthesis technology, research on prosody pattern generation techniques has been conducted in order to improve the naturalness of synthesized sounds. In addition, since the prosodic pattern determines the tone of the speech and is unique to the speaker, in speech recognition technology, for example, as shown in Non-Patent Document 1, etc. Research has been done on techniques for identifying meaning and speakers.

  In the field of speech synthesis technology, prosody parameters have been extracted from spontaneous speech waveforms and stored in a database, and research has been conducted on technology that uses the natural speech prosody parameters stored in the database during speech synthesis. I came. This method has an advantage that a prosodic pattern having high naturalness is used because the prosodic pattern is extracted from a speech waveform of a natural utterance. However, since it is physically impossible to memorize a pattern that matches all input texts, it is essential to transform the prosodic pattern, and this transformation process reduces the naturalness of the prosodic pattern. Problems arise. In order to solve such a problem, a technique for improving the naturalness of the prosodic pattern has been proposed.

  For example, in Patent Document 1, a representative pattern is generated for each cluster of prosodic patterns in the prosodic database in order to select a prosody pattern of a synthesized sound at the time of speech synthesis. A learning method for generating a representative pattern by evaluating an error from a prosodic pattern extracted from data is disclosed (see FIG. 13).

  Patent Document 2 discloses a method of creating a representative prosodic pattern by performing weighting according to the average or closeness of a plurality of prosodic patterns up to a certain number in order from a specific prosodic pattern.

  Further, Patent Document 3 discloses a method for calculating a prosodic parameter for a clause of input text data from a prosodic parameter relating to a clause having a similarity equal to or higher than a reference value compared to attribute information of the clause of input text data. (See FIG. 14). In this method, unlike the above-described conventional example, the similarity between the input text and the prosodic parameters in the speech corpus is calculated at the time of speech synthesis, so that a more prosodic pattern can be obtained.

By Sadahiro Furui, “Digital Speech Processing”, Tokai University Press, 1985 JP-A-11-95783 JP 2004-117663 A JP 2000-56788 A

  In the techniques shown in Patent Document 1 and Patent Document 2, a representative pattern created in advance is used for speech synthesis. For this reason, there is no choice but to select the prosody pattern of the synthesized sound from the patterns representing the category, and when the optimal prosodic pattern for the text actually input by the user is significantly different from the representative pattern created in advance. There was a problem that the naturalness and stability of the synthesized sound were impaired.

  In the technique shown in Patent Document 3, since the prosodic pattern is generated at the time of speech synthesis, the problem that the optimal prosodic pattern for the input text is greatly different is less likely to occur. In the calculation process, the similarity of attribute information between the input text data and the selected prosodic pattern is not used. For this reason, even if a pitch pattern whose attribute information is particularly similar to the input phonetic symbol string exists in the prosodic database, a pitch pattern similar to the pitch pattern is not generated, and a synthetic sound having high naturalness is obtained. There was a problem that it was difficult.

  Therefore, an object of the present invention is to provide a prosodic pattern generation device, a prosodic pattern generation method, and a prosodic pattern generation program that dynamically generate a prosody pattern having high naturalness with respect to an input text while maintaining high stability. Furthermore, it is more desirable to provide a prosodic pattern generation device, a prosodic pattern generation method, and a prosodic pattern generation program that can reduce the calculation load.

In order to achieve the above object, the prosody pattern generation apparatus according to the present invention includes a prosody database that stores in advance divided into categories in association with prosodic patterns and attribute information for each segment that is a constituent unit of a sentence;
Attribute information extraction means for extracting attribute information of the phonetic symbol string input;
Category selection means for specifying which category in the prosodic database divided into categories the input phonetic symbol string ;
The degree of similarity is determined according to the importance of the attribute information of the prosodic pattern existing in the prosodic database and the attribute information extracted from the input phonetic symbol string only for the category database specified by the category selecting means. Similarity calculation means for calculating;
Prosody pattern generation means for generating a new prosodic pattern by combining the prosodic patterns in the database of the category specified by the category selection means according to the weighting according to the similarity,
The weighting is performed by increasing the weight for the prosodic pattern having a large similarity and decreasing the weight for a prosodic pattern having a small similarity .

In this way, a prosodic pattern having higher naturalness while maintaining stability by generating an optimal prosodic pattern each time according to the attribute information of the input phonetic symbol string without using a representative pattern created in advance. Can be reproduced.
In particular, when generating a new prosodic pattern, the similarity is calculated according to the importance of the prosodic pattern attribute information existing in the prosodic database and the attribute information extracted from the input phonetic symbol string. Based on the degree, the prosodic patterns in the prosodic database are weighted and combined so that they are large for prosody patterns with high similarity and small for prosodic patterns with low similarity. If there is a prosodic pattern in the prosodic database that is particularly similar to the attribute information of the generated phonetic symbol string, a prosodic pattern that is particularly similar to that prosodic pattern is generated, and the input phonetic symbol string is highly natural. The prosodic pattern you have can be reproduced.
Even if there is no prosodic pattern with attribute information similar to phonetic symbol attribute information in the prosodic database, a prosodic pattern that averages multiple prosodic patterns in the prosodic database is generated, which is stable. Can be synthesized.
As described above, it is possible to create a synthesized sound having high naturalness for the input phonetic symbol string while maintaining high stability.

In particular , prosody pattern candidates used for combining when generating prosodic patterns can be limited to prosodic patterns within a category, so that the calculation load can be greatly reduced, improving processing speed and reducing storage capacity. Connected.

  The speech synthesizer of the present invention has a main part in common with the prosody pattern generation device described above, and further, waveform generation means for generating a speech waveform by controlling the prosody by the prosodic pattern generated by the prosody pattern generation means Have

  This provides a speech synthesizer capable of creating a synthesized sound having high naturalness for an input phonetic symbol string while maintaining high stability.

In order to achieve the same object as described above, the prosody pattern generation method of the present invention extracts the attribute information of the input phonetic symbol string;
A determination step of determining which category in the prosodic database into which the input phonetic symbol string is pre-categorized ;
Calculating similarity based on attribute information for each prosodic pattern stored in advance in the prosodic database of the category specified in the determining step and importance of attribute information extracted from the input phonetic symbol string When,
Generating a new prosodic pattern by combining prosodic patterns in the prosodic database of the category specified in the determining step according to weighting according to the similarity, and
The weighting in the step of generating the new prosodic pattern is performed by increasing the weight for the prosodic pattern having a large similarity and decreasing the weight for a prosodic pattern having a small similarity. .

An optimal prosodic pattern is generated each time according to the input phonetic symbol string attribute information without using a representative pattern created in advance, so the prosody pattern has higher naturalness while maintaining stability. The similarity is calculated according to the importance of the attribute information extracted from the input phonetic symbol string and the attribute information of the prosodic pattern existing in the prosody database. The prosodic pattern is weighted so that it is large for prosody patterns with high similarity and small for prosody patterns with low similarity, so that the prosodic patterns are combined. If there is a prosodic pattern in the prosodic database that is particularly similar to the attribute information of the input phonetic symbol string, It is possible to reproduce the prosody pattern with high naturalness, especially for similar product to enter the pronunciation symbol string prosodic patterns. Even if there is no prosodic pattern with attribute information similar to phonetic symbol attribute information in the prosodic database, a prosodic pattern that averages multiple prosodic patterns in the prosodic database is generated, which is stable. Can be synthesized. Therefore, it is possible to create a synthesized sound having high naturalness for the input phonetic symbol string while maintaining high stability.

In particular , prosody pattern candidates used for combining when generating a prosodic pattern can be limited to prosodic patterns within a category, so that the calculation load can be greatly reduced, the processing speed is improved, and the storage capacity is increased. It leads to reduction.

  The speech synthesis method of the present invention has a main part in common with the prosody pattern generation method described above, and further includes a step of generating a speech waveform by controlling the prosody using the generated prosody pattern.

  According to this speech synthesis method, it is possible to create a synthesized sound having high naturalness for the input phonetic symbol string while maintaining high stability.

In order to achieve the same object as described above, the prosody pattern generation program of the present invention includes a computer constituting the prosody pattern generation apparatus,
A process of extracting attribute information of the phonetic symbol string input;
A determination process for determining which category in the prosodic database in which the input phonetic symbol string is pre-categorized ;
Processing for calculating similarity according to the importance of attribute information for each prosodic pattern stored in advance in the prosodic database of the category specified in the determination processing and attribute information extracted from the input phonetic symbol string When,
Depending on the similarity, weighting is performed so that the prosodic pattern having a high similarity is large and the prosodic pattern having a low similarity is small, and the weight is set in the database of the category specified by the determination process . And a process for generating a new prosodic pattern by combining the prosodic patterns.

  A computer in which this prosodic pattern generation program is installed functions as the aforementioned prosodic pattern generation apparatus.

  The speech synthesis program of the present invention has a main part in common with the prosody pattern generation program described above, and further generates a speech waveform by controlling the prosody by the prosody pattern generated by the prosody pattern generation means in the computer. The function to perform is given.

  A computer installed with this speech synthesis program functions as the speech synthesis apparatus described above.

  The prosody pattern generation apparatus, prosody pattern generation method, and prosody pattern generation program according to the present invention generate an optimal prosody pattern each time according to attribute information of an input phonetic symbol string without using a representative pattern created in advance. As a result, prosodic patterns with higher naturalness can be reproduced while maintaining stability.

In particular, when generating a new prosodic pattern, the similarity is calculated according to the importance of the prosodic pattern attribute information existing in the prosodic database and the attribute information extracted from the input phonetic symbol string. Based on the similarity , the prosodic patterns in the prosodic database are weighted so that they are large for prosody patterns with high similarity and small for prosody patterns with low similarity. Therefore, if a prosodic pattern that is particularly similar to the attribute information of the input phonetic symbol string exists in the prosodic database, a prosodic pattern that is particularly similar to that prosodic pattern is generated and It can reproduce the prosody pattern with high naturalness, further category candidates prosodic pattern used for coupling in generating the prosody pattern Since the limited prosodic pattern of the inner can significantly reduce the computational load, leading to reduction of the increase and the storage capacity of the processing speed.

  Even if there is no prosodic pattern with attribute information similar to phonetic symbol attribute information in the prosodic database, a prosodic pattern that averages multiple prosodic patterns in the prosodic database is generated, which is stable. Can be synthesized.

  Therefore, it is possible to create a synthesized sound having high naturalness for the input phonetic symbol string while maintaining high stability.

  Next, the best mode for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an outline of the configuration of a computer 1 that functions as a speech synthesizer by installing a speech synthesis program for realizing the speech synthesis method of the present invention.

  This speech synthesis program includes a prosody pattern generation program to which the present invention is applied as a main part, and as a result, the computer 1 in which the speech synthesis program is installed applied the prosody pattern generation method of the present invention. It also functions as a prosodic pattern generation device.

  The computer 1 is composed of a normal workstation or personal computer, and includes a microprocessor (hereinafter simply referred to as CPU) 2 as a calculation means, a ROM 3 storing a basic control program for the CPU 2, and a temporary storage of calculation data. And the like, a hard disk 5 as a large-capacity storage device, and an interface 6 for connecting to various external devices and networks.

  A speech synthesis program is installed in the hard disk 5, and this speech synthesis program is expanded on the RAM 4 as necessary, and the CPU 2 that is driven and controlled according to the speech synthesis program has attribute information extraction means, similarity calculation means, prosody. Functions such as pattern generation means and waveform generation means are realized. Further, it is assumed that the hard disk 5 stores a prosodic database in advance.

  Further, a keyboard 7 and a monitor 8 functioning as a man-machine interface are connected to the CPU 2 via an input / output circuit 9, and a speaker as an audio output means is connected to the input / output circuit 9 via a driver 10. 11 is connected.

  FIG. 2 is a functional block diagram showing an outline of the functions of the CPU 2 that is driven and controlled by the speech synthesis program. The main parts thereof are attribute information extraction means 12, similarity calculation means 13, prosody pattern generation means 14, and It is comprised by the waveform generation means 15.

The attribute information extraction unit 12 analyzes a phonetic symbol string input from an external device on the RAM 4 or the hard disk 5 or via the interface 6 and extracts attribute information for each segment.
The phonetic symbol string is mainly composed of reading information, and further includes at least attribute information such as accent position information and sentence segment delimiter information.
The segment here refers to a linguistic or acoustic delimiter in a sentence such as a clause or an accent phrase, and is a basic constituent unit of a sentence.
The attribute information means the above-described accent position information, language information, and language or acoustic parameters such as time length information.
The prosody database 16 of the hard disk 5 stores many parameters of prosodic patterns and attribute information extracted from speech of natural utterances such as human real voice and their corresponding relationships in segment units.

The similarity calculation means 13 is a feature amount space between the attribute information of each segment of the input phonetic symbol string extracted by the attribute information extraction means 12 and the attribute information of the prosodic pattern of each segment existing in the prosodic database 16. Calculate internal distance.
The feature amount is quantified so that the similarity between attribute information can be measured as a distance. Hereinafter, the similarity between attribute information is expressed as a distance in the feature amount space. If this distance is small, the similarity is large, and if the distance is large, the similarity is small.

  The prosodic pattern generation means 14 generates a new prosodic pattern by combining the prosodic patterns in the prosodic database 16 with weighting according to the similarity, using the distance in the feature amount space obtained by the similarity calculation means 13. . As a method of combining prosodic patterns, for example, a method of adding prosodic patterns in the prosodic database 16 with weighting according to the distance in the feature space can be used.

  The waveform generation unit 15 generates a synthesized sound using the new prosody pattern generated by the prosody pattern generation unit 14.

  The operation of the best mode for carrying out the present invention will be described with reference to FIGS.

  First, a phonetic symbol string of a synthesized sound that the user wants to create is input to the attribute information extraction means 12 (step a1 in FIG. 3).

  Next, the attribute information extraction means 12 analyzes the input phonetic symbol string and extracts attribute information for each segment (step a2).

  Then, a distance in the feature amount space between the extracted attribute information and the attribute information of each prosodic pattern in the prosodic database 16 is obtained by the similarity calculation means 13 (step a3).

  Next, the prosodic pattern generation means 14 weights and combines the prosodic patterns in the prosodic database 16 using the distance in the feature amount space obtained by the similarity calculation means 13 to generate a new prosodic pattern (step) a4).

  Finally, the waveform generation unit 15 controls the prosody based on the new prosody pattern created by the prosody pattern generation unit 14 to generate a synthesized sound (step a5).

  In this embodiment, since the prosodic pattern for controlling the prosody of the synthesized sound is dynamically generated every time the phonetic symbol string is input, it is possible to generate an optimal prosodic pattern for the input phonetic symbol string.

  Further, since the prosodic patterns of the natural utterance in the prosodic database 16 are weighted and combined, the prosodic pattern having the same or very similar utterance content is present in the prosodic database 16 with the same phonetic symbol string and attribute information. The prosodic pattern that strongly reflects the characteristics of this prosodic pattern, in other words, a prosodic pattern that is very suitable for the input phonetic symbol string, is created, while the prosodic pattern that closely resembles the input phonetic symbol string is created in the prosody database. If not, the prosodic pattern is obtained by averaging features of a plurality of prosodic patterns having attribute information relatively similar to the input phonetic symbol string within the range registered in the prosodic database 16. Will be generated. Therefore, it is possible to realize a speech synthesizer that ensures both high naturalness and high stability of the prosodic pattern.

FIG. 4 shows a functional block diagram of another embodiment. In this embodiment, the CPU 2 functions not only as the attribute information extraction means 12, similarity calculation means 13, prosody pattern generation means 14, and waveform generation means 15 described above, but also as a category selection means 17. Yes.
In addition, in the hard disk 5, a prosodic database 18 divided into categories is registered instead of the prosodic database 16 described above.

  The category-divided prosody database 18 stores a prosody database equivalent to the prosody database 16 described above in advance by category division. As a category division method, there is a method of dividing by attribute information such as the number of mora and accent type.

  The category selection means 17 determines which category in the prosodic database 18 in which each segment of the input phonetic symbol string is divided into categories based on the attribute information extracted by the attribute information extraction means 12 from the input phonetic symbol string. And the corresponding category database is specified.

  The similarity calculation means 13 then applies only the database of the category specified by the category selection means 17 to the attribute information of each segment of the input phonetic symbol string and the prosodic pattern for each segment existing in the prosodic database 18. The distance in the feature amount space between the attribute information is calculated. The prosodic pattern generation means 14 performs weighting according to the calculation result of the similarity calculation means 13 and generates a prosodic pattern by performing a combining operation using only the prosodic patterns belonging to the category selected by the category selection means 17. .

  The operation of this embodiment will be described with reference to FIG. However, since the processing of step b1 to step b2 and step b4 to step b6 is the same as the processing of step a1 to step a2 and step a3 to step a5 (see FIG. 3) of the above-described embodiment, they overlap. Description of the portion is omitted, and only the operation of step b3 related to the newly added configuration will be described here.

  In the process of step b3, the category selection means 17 determines which category in the prosodic database 18 into which the phonetic symbol string is divided based on the attribute information extracted from the input phonetic symbol string. . Subsequent steps b4 and b5 are performed only in the category selected by the discrimination processing of the category selection means 17.

  In this embodiment, since a category that is a set of prosodic patterns having similar features is selected in advance before generating a new prosodic pattern, a certain degree of similarity is set when a new prosodic pattern is generated. It is possible to perform calculations using only the prosodic patterns that are possessed. For this reason, it becomes easy to improve the processing speed and obtain a stable prosodic pattern.

  FIG. 6 is a block diagram of a speech synthesis apparatus of an example corresponding to the embodiment shown in FIGS. The configuration of the computer 1 operating as a speech synthesizer and the functions of the CPU 2 are as described above. Here, the procedure relating to the speech synthesis method will be described exclusively from the method side.

  In this embodiment, it is assumed that the prosodic database 16 in which prosodic patterns are divided and stored in segment units is created in advance. It is assumed that one accent phrase in the sentence is one segment as the segment, and the prosody pattern is registered with the time series of the fundamental frequency normalized in the time direction and the frequency direction. The prosody database 16 stores at least attribute information such as at least the number of accent phrase mora, accent type, position in the sentence, phoneme string, and accent type of the preceding accent phrase for each prosody pattern. To do.

  Here, the prosodic pattern and the attribute information will be briefly described with reference to FIG. As a specific example, a sentence “synthesize speech” (tentatively called example sentence S) is handled. When the example sentence S is divided into accent phrases, it is divided into two accent phrases, “speech” (first accent phrase) and “synthesize” (second accent phrase).

  Further, when the example sentence S is read and converted into a hiragana string of information, it becomes “on's name / goose-shima's” (where 'is an accent symbol and / is an accent phrase delimiter).

  Here, when a part of the attribute information of the first accent phrase is extracted, the number of mora is 5, the accent type is 1, the position in the sentence is the beginning of the sentence, and the attribute information of the second accent phrase is the mora number of 7, the accent type However, the position in the sentence is the end of the sentence.

  The position in the sentence is divided into a sentence head, a sentence end, a sentence, a word, etc., and each is digitized as a feature value representing the position in the sentence.

  In the series of processes in this embodiment, paying attention to the nth accent phrase (referred to as the “nth accent phrase”) when a phonetic symbol string consisting of N accent phrases is input. explain. In this phonetic symbol string, at least an accent position and an accent phrase break are specified as attribute information.

とする。但し、Ｊは属性情報の種類の数であり、例えば、図６に示される例のように、属性情報がモーラ数，アクセント型，文中の位置，先行アクセント型の４種であれば、Ｊの値は４となる。また、ｊは属性情報の種類のインデックスであり、この例では、ｊ＝１がモーラ数，ｊ＝２がアクセント型，ｊ＝３が文中の位置，ｊ＝４が先行アクセント型を意味する。 First, attribute information extraction processing is performed on the nth accent phrase of the input phonetic symbol string as shown in FIG. 6 (step of extracting attribute information). The types of attribute information to be extracted include at least those stored for the prosodic patterns in the prosodic database 16. The value of this attribute information

And However, J is the number of types of attribute information. For example, as in the example shown in FIG. 6, if the attribute information has four types of mora number, accent type, position in sentence, and preceding accent type, The value is 4. Further, j is an index of the type of attribute information. In this example, j = 1 indicates the number of mora, j = 2 indicates the accent type, j = 3 indicates the position in the sentence, and j = 4 indicates the preceding accent type.

For each type of attribute information, when calculating the distance in the feature amount space, for example, if the number of mora and accent types that are highly important attribute information do not match, the distance increases (similar The weight α _j is set so that the degree becomes smaller. The weight α _j is set so as to satisfy the following expression (1), that is, the total sum of the weights is 1.

Next, one piece of attribute information stored corresponding to each segment of the prosodic pattern in the prosodic database 16 is read as shown in FIG. The value of this attribute information is assumed to be a _ij (i = 1, 2, 3,..., I, j = 1, 2, 3,..., J). Here, I is the total number of prosodic patterns of segments registered in the prosodic database 16, and i is an index thereof.

Next, the relative distance in the feature quantity space is calculated as shown in FIG. 6 between the input attribute information of the phonetic symbol string and the attribute information of the prosodic pattern read from the prosodic database 16 (similarity). Step to calculate the degree).
The relative distance in the feature amount space is obtained by adding the weighted difference between the parameters of the input phonetic symbol string attribute information and the attribute information read out from the prosodic database 16 to obtain the number of mora of each accent phrase. Divide by to get it.

但し、Ｍは入力された発音記号列のセグメント内のモーラ数である。 Therefore, the relative distance d _i in the feature amount space between the input phonetic symbol string and the prosodic pattern of the index i in the prosodic database 16 is expressed by the following equation (2).

However, M is the number of mora in the segment of the inputted phonetic symbol string.

ここで、Ｄは距離の総和、つまり、韻律データベース１６内に記憶されたｉ＝１〜Ｉの韻律パターンの各々について求めた特徴量空間内相対距離ｄ_ｉを全て加算した値である。 Next, the weight w _i for combining the prosodic patterns is calculated using the relative distance d _i in the feature amount space thus obtained. The weight w _i is obtained by the following equation (3).

Here, D is a sum of distances, that is, a value obtained by adding all the relative distances d _i in the feature amount space obtained for each of the prosodic patterns of i = 1 to I stored in the prosodic database 16.

Next, using the weight w _i, and generates a new prosodic pattern by multiplying the weights w _i corresponding to each prosodic pattern of the stored index i in the prosodic database 16 performs weighting by linear combination ( Generating a new prosodic pattern).

  The same processing is repeated for the nth accent phrase of n = 1 to N, and the prosody of the speech waveform is controlled for each accent phrase by the new prosodic pattern generated corresponding to each accent phrase. Further, the segments are connected by correcting the time length and the like, and finally a synthesized sound of one whole sentence composed of N accent phrases is generated (step of generating a speech waveform).

  As a specific example, the present embodiment will be described with reference to an example sentence S. An outline of this embodiment is shown in FIG. FIG. 8A visually shows the characteristics (relationship between normalized fundamental frequency and time series) of individual prosodic patterns registered in the prosodic database 16, and FIG. The position of the prosodic pattern in the feature amount space, that is, information related to the relative similarity of the attribute information registered corresponding to each prosodic pattern is visualized and shown.

  In FIG. 8B, points A and B are positions in the feature amount space of the first accent phrase and the second accent phrase, respectively. Each pattern of A1 to A3, B1 to B3, C, and D shown in FIG. 8A represents some of the prosodic patterns stored in the prosodic database 16, This corresponds to A1 to A3 and B1 to B3 shown in the feature space diagram in FIG.

Therefore, first, as shown in FIG. 8B, the feature amount between the point A corresponding to the attribute information of the first accent phrase and the attribute information of each prosodic pattern stored in the prosodic database 16 space relative distance _{d AA1, d AA2, d AA3} , d AB1, d AB2, d AB3, d AC, d AD,. . . , The weight _{w AA1, w AA2, w AA3} , w AB1, w AB2, w AB3, w AC, w AD of each prosodic patterns used to generate a new prosodic pattern according to the calculated, equation (2). . . To decide.

As for the point B corresponding to the attribute information of the second accent, similar to the _{above, d BA1, d BA2, d} BA3, d BB1, d BB2, d BB3, d BC, d BD,. . . And weights w _BA1 , w _BA2 , w _BA3 , w _BB1 , w _BB2 , w _BB3 , w _BC , w _{BD ,.} . . To decide.

At this time, as shown in FIG. 8 (b), the distance between A-A1, A-A2, and A-A3 is shorter than the other points, so it corresponds to these three points A1, A2, A3. Prosody pattern weights w _AA1 , w _AA2 , and w _AA3 are larger than others, and the distance between A-A1 is particularly short, resulting in a prosody very similar to the prosody pattern indicated by point A1. The weighting to be a pattern is determined.

  Similarly, since the distance between B-B1, B-B2, and B-B3 is shorter than other points, a prosodic pattern obtained by averaging the prosodic patterns corresponding to these three points B1, B2, and B3 is obtained. The weights to be determined are determined. In this case, no prosodic pattern having attribute information remarkably similar to the input prosodic pattern attribute information exists in the prosodic database 16.

Finally, the weights obtained _{w AA1, w AA2, w AA3} , w AB1, w AB2, w AB3, w AC, w AD, first accent phrase of the sentence S by linear combination weighted with ... And a new prosodic pattern related to the generated weights w _BA1 , w _BA2 , w _BA3 , w _BB1 , w _BB2 , w _BB3 , w _BC , w _{BD ,.} . . Is used to generate a new accent phrase prosodic pattern related to the second accent phrase of the example sentence S.

  Next, the processing operations of the CPU 2 functioning as the attribute information extraction means 12, the similarity calculation means 13, the prosody pattern generation means 14, and the waveform generation means 15 are outlined in FIG. A specific description will be given from the aspect of internal processing of the CPU 2 with reference to the flowchart.

  First, the CPU 2 starts reading a phonetic symbol string on the RAM 4, the hard disk 5, or an external device (step c1), and divides the read phonetic symbol string into accent phrases in order from the top (step c2).

  Therefore, according to the above example, the n = 1st accent phrase of a certain sentence consisting of N accent phrases is read first.

Next, the CPU 2 functioning as the attribute information extraction unit 12 extracts each attribute information a _j of j = 1 to J from this accent phrase (step c3).

According to the above-described example, j = 1 attribute information a ₁ is the number of mora, j = 2 attribute information a ₂ is an accent type, j = 3 attribute information a ₃ is a position in the sentence, and j = 4 attribute information. a ₄ is a preceding accent type, comprising the attribute information of all four are extracted.

  Next, the CPU 2 initializes the distance integrated value register D for obtaining the sum D of distances in the above-described equation (3) to 0 (step c4), and further specifies the data in the prosodic database 16 to be read. Is set once to the prosodic pattern identification index i in the database (step c5), the index i is immediately incremented by 1, and the data corresponding to the first prosodic pattern registered in the prosodic database 16 is obtained. Update to the initial value 1 for reading (step c6).

Next, the CPU 2 corresponds to the value of Σα _j | a _j −a _ij | in the above formula (2), that is, one attribute information a _{j included in} the input phonetic symbol string and the prosody database 16 corresponding to this. A value obtained by multiplying the difference from one attribute information a _{ij included in} the prosodic pattern of index i read out from the above by a predetermined weighting coefficient α _j is added to all the attribute information of j = 1 to J. The value of the integrated value storage register X for matching is initialized to 0 (step c7), and further, an attribute information specifying index for specifying the type of attribute information included in the prosodic pattern of index i read from the prosodic database 16 A value that represents the first attribute information of each prosodic pattern in the prosodic database 16 by immediately incrementing the index j after setting j to 0 once (step c8). , That is, in this example updates the value 1 representing the number of moras (Step c9).

Next, the CPU 2 reads out the jth attribute information a _ij of the prosodic pattern of the index i from the prosodic database 16 based on the current value of the prosodic pattern specifying index i in the database and the current value of the attribute information specifying index j (step c10), the value of α _j | a _j −a _ij | in the above equation (2) is obtained (step c11), and this value is added to the integrated value storage register X (step c12).

Therefore, at the present time when i = 1 and j = 1, the first attribute information a _{11 included in} the prosodic pattern of index 1 from the prosodic database 16, that is, the number of mora a ₁₁ of the prosodic pattern first registered in the prosodic database 16 is obtained. A difference from the _first attribute information a _1, that is, the number of mora included in the phonetic symbol string read out and input is obtained, and this difference is multiplied by a weighting coefficient α ₁ (set value) corresponding to the number of mora. The value obtained in this way is added to the integrated value storage register X.

Next, the CPU 2 determines whether or not the current value of the attribute information identification index j has reached the total number J of attribute information types, that is, for all attribute information of j = 1 to J included in the prosodic pattern of the index i. It is determined whether or not the process of finding the difference from the attribute information of the input phonetic symbol string corresponding to is multiplied by the weighting coefficient α _j is completed (step c13).

  If the determination result in step c13 is true, that is, if the processing regarding all attribute information of j = 1 to J included in the prosodic pattern of index i has not been completed, the CPU 2 The same processing as described above is repeatedly executed while incrementing the value of the specific index j by 1 (step c9 to step c13).

Finally, when the determination result in step c13 is false and the value of the attribute information identification index j reaches the total number J of attribute information types, Σα _j | a _j −a _ij in equation (2) The value of | is obtained by the integrated value storage register X.

Therefore, the CPU 2 functioning as the similarity calculation means 13 is the value of the accumulated value storage register X, that is, the value of Σα _j | a _j −a _ij | in the equation (2) when the determination result in step c13 becomes false. and in the segment of the input string of phonetic symbols divided by number of moras M, the value of d _i in the equation (2), i.e., the prosody pattern of the index i in the string of phonetic symbols and prosody database 16 that is input A relative distance d _i in the feature amount space is obtained (step c14).

Next, the CPU 2 adds the value of the relative distance d _i in the feature amount space obtained this time to the distance integrated value register D (step c15), and the current value of the prosodic pattern specifying index i in the database is stored in the prosodic database 16. Whether or not the total number I of the prosodic patterns registered has been reached, in short, the value of the relative distance d _i in the feature amount space for all the prosodic patterns of indexes i = 1 to I registered in the prosodic database 16 It is determined whether or not it has been obtained (step c16).

If the determination result in step c16 is true, that is, if it is determined that the prosodic pattern for which the value of the relative distance d _i in the feature amount space is not found remains in the prosodic database 16, The CPU 2 repeatedly executes the same processing as described above while incrementing the value of the prosodic pattern identification index i in the database by 1 (step c6 to step c16).

Then, when the determination result in step c16 is finally false, the value of the relative distance d _i in the feature amount space for all the prosodic patterns of indexes i = 1 to I registered in the prosodic database 16 is obtained. At the same time, the distance sum value D in the above equation (3) is obtained by the distance integrated value register D.

Therefore, the CPU 2 functioning as the prosodic pattern generation means 14 has the weights w _i = D in the above-described equation (3) for all the prosodic patterns of indexes i = 1 to I when the determination result in step c16 is false. / D _i is obtained individually, and all prosodic patterns with indices i = 1 to I are subjected to linear combination processing by weighting with w _i to generate a new prosodic pattern for the one accent phrase, The contents are temporarily stored in the RAM 4 (step c17).

In short, this processing is such that each of the prosodic patterns of indexes i = 1 to I in the prosodic database 16 is multiplied by the corresponding weight w _i and added over i = 1 to I.

  Thus, when the generation of a new prosodic pattern for one divided accent phrase is completed, the CPU 2 determines whether or not the generation of the prosodic pattern has been completed for all of the divided accent phrases (step c18) If generation of prosodic patterns for all accent phrases has not been completed, the CPU 2 returns to the process of step c2 again to divide the next accent phrase from the above-mentioned one sentence, and for this accent phrase By repeating the same process as described above, a prosodic pattern corresponding to a newly divided accent phrase is generated (step c2 to step c18).

  When the determination result in step c18 is finally true and new prosodic patterns are generated for all of the n = 1st to Nth accent phrases of one sentence composed of N accent phrases, the waveform generating means The CPU 2 functioning as 15 reads out N new prosodic patterns corresponding to each of the n = 1 to Nth accent phrases from the RAM 4, corrects the time length, etc., and connects these segments (step c19) Finally, a synthesized sound of one whole sentence composed of N accent phrases is generated (step c20).

  According to this embodiment, when a prosodic pattern very similar to the attribute information of the input phonetic symbol string exists in the prosodic database 16, a prosodic pattern very close to the prosodic pattern extracted from the natural utterance is generated. Therefore, a synthesized sound having very high naturalness is generated.

  Even if there is no prosodic pattern having similar attribute information in the prosodic database 16, a prosodic pattern that is an average of a plurality of prosodic patterns in the prosodic database 16 is generated. Synthesis can be performed.

In this embodiment, the relative distance in the feature amount space of the attribute information of each prosodic pattern and the attribute information of the input phonetic symbol string is calculated based on each attribute information. A reference origin is defined, and the distance between the origin and each prosodic pattern in the feature amount space is calculated in advance and stored in the prosodic database 16, and between the input phonetic symbol string attribute information and the origin There is also a method of obtaining the distance d _i in the feature amount space by the difference between this distance and the distance from the origin in the attribute information of each prosodic pattern. By this method, the number of distance calculations when generating a new prosodic pattern can be reduced, and the calculation time can be further reduced.

  In this embodiment, the phonetic symbol string including the accent phrase delimiter and the accent position is input. However, linguistic information such as part-of-speech information, dependency information, and sending kana information may also be included. Is possible. By including linguistic information, feature quantities that cannot be extracted from acoustic information alone or difficult to extract can be used as parameters for prosodic pattern generation. Of course, linguistic information can be given to prosodic patterns in the prosodic database.

  Next, an example corresponding to the embodiment shown in FIGS. 4 and 5 will be briefly described with reference to the drawings.

  FIG. 10 is a block diagram of a speech synthesizer according to an example corresponding to the embodiment shown in FIGS. In addition to the configuration of the first embodiment described above, the present embodiment includes a category selection unit 17 that selects a category in the prosodic database 18 based on the relative distance in the feature amount space obtained by the similarity calculation unit 13. The configuration of the computer 1 operating as a speech synthesizer and the functions of the CPU 2 are as described above. Here, the procedure relating to the speech synthesis method will be described exclusively from the method side.

  Now, it is assumed that a phonetic symbol string is input and attribute information is extracted from the input phonetic symbol string in the same manner as described in the first embodiment.

  The category selection means 17 determines that the attribute information extracted from the input phonetic symbol string belongs to the category division attribute information in the prosodic database 18 and belongs to that category, and at the time of prosodic pattern generation A new prosodic pattern is generated using the prosodic pattern belonging to the selected category (step of determining which category in the prosodic database into which the input phonetic symbol string is divided into categories).

  The above example sentence S will be described in detail as an example. As described above, information such as the number of mora is 5, the accent type is 1, and the position in the sentence is the beginning of the sentence is extracted from “speech” of the first accent phrase of the example sentence S.

  On the other hand, as shown in FIG. 11, if category 1 in the prosodic database 18 is a category to which a prosodic pattern having information that the number of mora = 5, the accent type = 1, and the position in the sentence = the beginning of the sentence belongs, The category selection means 17 determines that the first accent phrase of S belongs to category 1.

  Similarly, if category 2 in the prosodic database is a category to which the prosodic pattern has information that the number of mora = 7, accent type = 6, position in sentence = sentence end, the second accent phrase "synthesize" The category selection means 17 determines that the image belongs to category 2.

  Here, as a category dividing method, for example, a category having a certain range such as “the number of mora is 6 or more and the accent type is 5 or 6” may be used.

ここで、ｄ_ｉ’は選択されたカテゴリに属する韻律パターンの特徴量空間内相対距離、また、Ｄ'は選択されたカテゴリに属する韻律パターンの特徴量空間内相対距離の合計、そして、Ｉ'は選択されたカテゴリに属する韻律パターンの総数である。 When a category is selected in this way, next, the weight of each prosodic pattern belonging to that category is calculated. The weight w _i ′ for each prosodic pattern belonging to the selected category is expressed by the following equation (4).

Here, d _i ′ is the relative distance in the feature amount space of the prosodic pattern belonging to the selected category, D ′ is the sum of the relative distances in the feature amount space of the prosodic pattern belonging to the selected category, and I ′ Is the total number of prosodic patterns belonging to the selected category.

Finally, the weight w _i ′ is used, and each of the prosodic patterns of indexes i = 1 to I ′ belonging to the selected category is multiplied by the corresponding weight w _i ′ to perform linear combination. As in the case of 1, a new prosodic pattern to be obtained is generated.

  As a specific example, the present embodiment will be described with reference to an example sentence S. An outline of this embodiment is shown in FIG. In FIG. 11, the same reference numerals as those in FIG. 8 denote the same or corresponding parts, and the description thereof will be omitted.

  The prosodic patterns in the prosodic database 18 are divided into a number of categories. In FIG. 11, the prosody of the category 1 to which the prosodic pattern of the number of mora = 5, the accent type = 1, the number of mora = 7, and the accent type = 6. Only category 2 to which the pattern belongs is shown.

  As described above, it is assumed that point A and point B belong to category 1 and category 2, respectively. A1, A2, and A3 are included in category 1, and B1, B2, and B3 are included in category 2. Shall belong.

First, for the first accent phrase A, the prosodic patterns B1, B2, B3 and C, D that are out of the category 1 are ignored, and the relative in-feature space with the prosodic patterns A1, A2, A3 belonging to the category 1 is ignored. Only the distances d _AA1 , d _AA2 , and d _AA3 are calculated, and the weights w _AA1 , w _AA2 , and w _AA3 are determined according to the equation (4).

Next, with respect to B of the second accent phrase, prosodic patterns A1, A2, A3 and C, D out of category 2 are ignored, and within the feature amount space between prosodic patterns B1, B2, B3 belonging to category 2 Only the relative distances d _BB1 , d _BB2 , and d _BB3 are calculated, and the weights w _BB1 , w _BB2 , and w _BB3 are determined according to Equation (4).

  At this time, as in the case of the first embodiment, for the first accent phrase A, the prosody pattern very similar to A1, and for the second accent phrase B, the prosody pattern that averages B1, B2, and B3. The weight is generated so that is generated.

Finally, using the weights w _AA1 , w _AA2 , w _AA3 and w _BB1 , w _BB2 , w _BB3 obtained in this way, the first accent phrase and the second accent phrase of the example sentence S by weighted linear combination Generate a new prosodic pattern for.

  Next, the outline of the speech synthesis program installed in the hard disk 5 with respect to the processing operations of the CPU 2 functioning as the attribute information extraction means 12, similarity calculation means 13, prosodic pattern generation means 14, waveform generation means 15, and category selection means 17. The internal processing of the CPU 2 will be specifically described with reference to the flowchart of FIG.

The processing from step d1 to step d3 is the same as the processing from step c1 to step c3 in FIG. 9. By these processing, first, the attribute information extraction means 12 is targeted for one accent phrase of the read phonetic symbol string. The CPU 2 that functions as _j extracts attribute information a _j for j = 1 to J.

  Next, the CPU 2 once sets 0 in a category designation index k for designating a category in the prosodic database 18 (step d4), and then immediately increments the index k by 1 to set the first category set in the prosodic database 18 Is updated to the initial value 1 for designating (step d5).

  Next, the CPU 2 functioning as the category selection means 17 reads out the type of attribute information registered as the category k from the prosodic database 18 based on the current value of the category designation index k (step d6), and for the category k It is determined whether or not the set attribute information matches the accent phrase attribute information of the phonetic symbol string read in the current process at a rate equal to or greater than the set value (step d7).

  The purpose of the category selection processing in this embodiment is to reduce the number of prosodic patterns that are actually subject to arithmetic processing and increase the efficiency of processing operations. The number of attribute information registered in a category can be arbitrarily determined. However, in order to avoid flooding various categories in the prosodic database 18, the total number of attribute information registered for each category is usually larger than the number J of accent phrases extracted from the accent phrase of the phonetic symbol string. Less. That is, the set value used in step d7 is a number smaller than J.

  Here, when the determination result in step d7 is true, the CPU 2 functioning as the category selection means 17 considers that the accent phrase of the phonetic symbol string read in this processing belongs to the category k, and If the determination result in step d7 is false, it is assumed that the accent phrase of the phonetic symbol string read in this processing does not belong to the category k.

  When the determination result in step d7 is false, that is, when it is determined that the accent phrase of the phonetic symbol string read in this processing does not belong to the category k, the function as the category selection unit 17 The CPU 2 that returns to the process of step d5 again increments the value of the category designation index k by 1 and repeatedly executes the same process as described above, so that the accent phrase of the phonetic symbol string read in this process is obtained. The value of category k that can be regarded as belonging is obtained.

  In this way, when the value of the category k that can be regarded as belonging to the accent phrase of the phonetic symbol string read in the current process is obtained, the CPU 2 functioning as the category selection means 17 at this time, The category to be read is limited to only category k (step d8), and the operations of the similarity calculation means 13 and the prosody pattern generation means 14 are allowed.

The processing in step d9 is the same as the processing in steps c4 to c17 in FIG. 9, and the CPU 2 functioning as the similarity calculation means 13 and the prosodic pattern generation means 14 performs the relative distance d in the feature amount space in the same manner as described above. The processing for obtaining _i ′, the weight w _i ′, etc. is repeatedly executed (however, d _i , w _i , I, D in FIG. 9 are d _i ′, w _i ′, I ′, D, respectively). It shall be read as').

However, in this embodiment, the category is limited by the processing of the above-described steps d4 to d8, that is, the function of the category selection means 17, and the target for obtaining the relative distance d _i ′ in the feature amount space, the weight w _i ′, etc. The number of prosodic patterns is not the total number I of all the prosodic patterns registered in the prosodic database 16, but the total number I ′ of prosodic patterns belonging to the selected category k. The load on the CPU 2 functioning as the means 13 and the prosodic pattern generation means 14 is greatly reduced, and the special effect of greatly improving the processing efficiency as a whole is achieved.

  The processing from step d10 to step d12 is the same as the processing from step c18 to step c20 in FIG.

  As described above, if the present embodiment is used, the prosodic patterns used for the weighted linear combination are limited, so that the stability and naturalness substantially equivalent to those of the first embodiment can be realized while reducing the calculation load. become able to.

It is the block diagram which showed the outline of the structure of the computer which functions as a speech synthesizer by installing the speech synthesis program for implement | achieving the speech synthesis method of this invention. FIG. 2 is a functional block diagram illustrating an outline of functions of a CPU that is driven and controlled by a speech synthesis program (first embodiment). It is the flowchart which simplified and showed the operation | movement of the embodiment. It is a functional block diagram of other embodiments showing an outline of functions of a CPU that is driven and controlled by a speech synthesis program (second embodiment). It is the flowchart which simplified and showed the operation | movement of the embodiment. FIG. 4 is a block diagram of an example speech synthesis apparatus corresponding to the embodiment shown in FIGS. 2 and 3 (Example 1). It is the conceptual diagram shown about the relationship between a prosodic pattern and attribute information (Example 1). FIG. 8A is a conceptual diagram showing the characteristics of individual prosodic patterns registered in the prosodic database, and FIG. 8B is a visualizing position of the prosodic patterns in the feature amount space. (Example 1) which is a conceptual diagram. 3 is a flowchart showing an outline of a speech synthesis program installed on a hard disk (Example 1). FIG. 6 is a block diagram of a speech synthesis apparatus of an example corresponding to the embodiment shown in FIGS. 4 and 5 (Example 2). (Example 2) which is the conceptual diagram which visualized and showed the characteristic of each prosodic pattern and the position of the prosodic pattern in the feature-value space which were divided into categories and registered in the prosodic database. 10 is a flowchart showing an overview of a speech synthesis program installed on a hard disk (Example 2). It is the conceptual diagram shown about the conventional prosodic pattern generation method which evaluates the difference | error of the deformation pattern which deform | transformed the representative pattern, and the prosodic pattern extracted from audio | voice data, and produces | generates a representative pattern. A conceptual diagram showing a conventional prosodic pattern generation method for calculating the prosodic parameters for the clauses of the input text data from the prosodic parameters for the clauses having a similarity greater than or equal to the reference value compared with the attribute information of the clauses of the input text data is there.

Explanation of symbols

1 Computer 2 CPU
3 ROM
4 RAM
5 hard disk 6 interface 7 keyboard 8 monitor 9 input / output circuit 10 driver 11 speaker 12 attribute information extraction means 13 similarity calculation means 14 prosody pattern generation means 15 waveform generation means 16 prosody database 17 category selection means 18 prosody database

Claims (6)

A prosodic database that stores a segment divided into categories in advance in association with prosodic patterns and attribute information for each segment that is a constituent unit of a sentence;
Attribute information extraction means for extracting attribute information of the phonetic symbol string input;
Category selection means for specifying which category in the prosodic database divided into categories the input phonetic symbol string ;
The degree of similarity is determined according to the importance of the attribute information of the prosodic pattern existing in the prosodic database and the attribute information extracted from the input phonetic symbol string only for the category database specified by the category selecting means. Similarity calculation means for calculating;
Prosody pattern generation means for generating a new prosodic pattern by combining the prosodic patterns in the database of the category specified by the category selection means according to the weighting according to the similarity,
The prosody pattern generation apparatus according to claim 1, wherein the weighting is performed by increasing the weight for a prosodic pattern having a large similarity and decreasing the weight for a prosodic pattern having a small similarity . A prosodic database that stores a segment divided into categories in advance in association with prosodic patterns and attribute information for each segment that is a constituent unit of a sentence;
Attribute information extraction means for extracting attribute information of the phonetic symbol string input;
Category selection means for determining which category in the prosodic database divided into categories the input phonetic symbol string ;
The degree of similarity is determined according to the importance of the attribute information of the prosodic pattern existing in the prosodic database and the attribute information extracted from the input phonetic symbol string only for the category database specified by the category selecting means. Similarity calculation means for calculating;
Prosody pattern generating means for generating a new prosodic pattern by combining prosodic patterns in the database of the category specified by the category selecting means according to weighting according to the similarity,
Waveform generating means for generating a speech waveform by controlling the prosody by the generated prosodic pattern,
The speech synthesizer according to claim 1, wherein the weighting is performed by increasing the weight for a prosodic pattern having a large similarity and decreasing the weight for a prosodic pattern having a small similarity . A prosodic pattern generation method for generating a prosodic pattern by a prosodic pattern generating device,
Extracting attribute information of the input phonetic symbol string;
A determination step of determining which category in the prosodic database into which the input phonetic symbol string is pre-categorized ;
Calculating similarity based on attribute information for each prosodic pattern stored in advance in the prosodic database of the category specified in the determining step and importance of attribute information extracted from the input phonetic symbol string When,
Generating a new prosodic pattern by combining prosodic patterns in the prosodic database of the category specified in the determining step according to weighting according to the similarity, and
Prosody pattern generation characterized in that weighting in the step of generating a new prosodic pattern is performed by decreasing the prosody pattern having a high similarity and decreasing the prosody pattern having a low similarity. Method. A speech synthesis method for generating synthesized speech by a speech synthesizer,
Extracting attribute information of the input phonetic symbol string;
A determination step of determining which category in the prosodic database into which the input phonetic symbol string is pre-categorized ;
Calculating similarity based on the attribute information of each prosodic pattern stored in advance in the database of the category specified in the determination step and the importance of the attribute information extracted from the input phonetic symbol string; ,
Generating a new prosodic pattern by combining prosodic patterns in the database of the category identified in the determining step according to weighting according to the similarity;
Generating a speech waveform by controlling the prosody according to the generated prosodic pattern,
The speech synthesis method characterized in that the weighting in the step of generating the new prosodic pattern is performed by increasing the weight for the prosodic pattern having a large similarity and decreasing the weight for a prosodic pattern having a small similarity. . In the computer constituting the prosody pattern generation device,
A process of extracting attribute information of the phonetic symbol string input;
A determination process for determining which category in the prosodic database in which the input phonetic symbol string is pre-categorized ;
Processing for calculating similarity according to the importance of attribute information for each prosodic pattern stored in advance in the prosodic database of the category specified in the determination processing and attribute information extracted from the input phonetic symbol string When,
Depending on the similarity, weighting is performed so that the prosodic pattern having a high similarity is large and the prosodic pattern having a low similarity is small, and the weight is set in the database of the category specified by the determination process . A prosody pattern generation program for executing a process of generating a new prosody pattern by combining the prosodic patterns. In the computer that composes the speech synthesizer,
A process of extracting attribute information of the phonetic symbol string input;
A determination process for determining which category in the prosodic database in which the input phonetic symbol string is pre-categorized ;
Processing for calculating similarity according to the importance of attribute information for each prosodic pattern stored in advance in the database of the category specified in the determination processing and attribute information extracted from the input phonetic symbol string; ,
In accordance with the similarity, weighting is performed so that the prosodic pattern having the high similarity is large and the prosodic pattern having the low similarity is small, and the weight is set in the database of the category specified by the determination process . Generating a new prosodic pattern by combining the prosodic patterns of
A speech synthesis program that executes a process of generating a speech waveform by controlling a prosody using the generated prosodic pattern.

Images