JP3286353B2 - Voice synthesis method - Google Patents

Abstract

PURPOSE:To improve the naturalness of a synthesized speech by generating a pitch pattern in consideration of the quantity of variation in the pitch frequency of the center of gravity between two adjacent morae. CONSTITUTION:This method consists of a pitch difference estimating process 9 for estimating the difference in pitch frequency of the center of gravity between two adjacent morae corresponding to parameter information on an accent type and parameter information on the number of morae, a pitch pattern generating process 8 for generating the pitch pattern of an optional character string according to the difference in pitch frequency, and a phoneme piece connecting process 10 for generating speech data by reading phoneme pieces required for speech vocalization out of a phoneme piece memory 11 stored with phoneme pieces consisting of speech parameters required for speech synthesis. In the pitch estimating process 9, the difference between a certain pitch frequency and the pitch frequency of the center of gravity of the mora if the mora is at the head of a sectioned range.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for generating a pitch pattern of a synthesized speech which greatly affects the phonology of speech and the naturalness of accents.

[0002]

2. Description of the Related Art In recent years, research on rule-based speech synthesis has been actively conducted as an important technology of human interface. Rule speech synthesis determines part-of-speech information by morphological analysis from character strings and the like, determines word reading by collation with a word dictionary, and then determines the accent type, accent combination, phrase, etc. of the word according to this reading, A pitch pattern is determined from these pieces of information, and a speech unit (for example, a Percoll coefficient or an LSP coefficient) corresponding to the reading of a word
Are connected to generate audio data.

[0003] That is, speech data is a Percoll coefficient sequence, a pitch pattern and amplifier information corresponding thereto, and among them, the pitch pattern is positioned as having a great effect on the naturalness of synthesized speech. .

In the conventional pitch pattern generation, after determining a phrase component and an accent component from a sentence structure,
It is known to use a point pitch model that generates a pitch pattern by estimating the pitch of the center of gravity of each mora and performing linear interpolation (see IEICE Transactions Vol.J63-D).
No. 9 pp. 715-722, 1980.9).

There is also known a method of estimating a value of each pitch frequency of a leading mora, a mora at which a pitch frequency takes a peak, and a trailing mora for each phrase by using a neural network (Speech Technical Committee Material SP89-111, 1990). .
1).

[0006] Each of these conventional methods for generating a pitch pattern is a method for unambiguously estimating the absolute value of the pitch frequency for each mora, and the connection between two adjacent mora (the amount of change in the pitch frequency) is not described. Not considered.

Therefore, in these methods, there is a problem that the amount of change in the pitch pattern between the mora is not stable, and a sense of discomfort is caused as compared with a case where a human talks.

[0008]

As described above, the factor that influences the naturalness of the rule-synthesized speech is that the pitch frequency of the center of gravity between two adjacent mora is not the absolute value of the pitch frequency of each mora. Although the amount of change is more important, the conventional method does not estimate the amount of change in pitch frequency between each mora.

SUMMARY OF THE INVENTION The present invention has been made to solve such a problem. In order to improve the naturalness of a synthesized speech in a ruled speech synthesis, a pitch frequency which greatly affects the naturalness of a synthesized speech is provided. The pitch pattern is to be generated in consideration of the amount of change in the pitch pattern.

[0010]

Means for Solving the Problems The present invention provides at least:
Correspondence table storing the difference in pitch frequency between each mora of a plurality of sentences composed of m accent phrases corresponding to the position of the accent phrase in the sentence, the parameter information of the accent type, and the parameter information of the number of mora. Alternatively, a neural network that has been learned to output accent type parameter information or mora number parameter information to the input layer and to output the difference in pitch frequency of the center of gravity between two adjacent mora from the output layer. , A pitch difference estimation processing unit for estimating a pitch frequency difference between two adjacent moras, and a pitch pattern generation unit for forming a pitch pattern based on the pitch frequency difference estimated by the pitch difference estimation processing unit And a processing unit.

When the target mora is located at the beginning of the divided range, the pitch difference estimation processing section assumes a constant pitch frequency as a preceding mora of the target mora, and sets a pitch of the center of gravity with the mora. In addition to estimating the frequency difference, if the accent phrase including the target mora is located at the beginning of the delimited range, the accent type of the preceding accent phrase of the accent phrase is undulated or flat. It is composed of a processing unit for estimating a difference in pitch frequency.

Further, the pitch difference estimation processing section estimates a pitch frequency difference between centroid points of two adjacent moras based on a pitch frequency difference between mora of a sentence composed of m accent phrases. In order to estimate the difference between the pitch frequencies of a sentence composed of n (> m) accent phrases, (m
Assuming that the position of the (+1) th to n-th accent phrases in the sentence is m, the processing unit estimates the difference in pitch frequency of the center of gravity between the mora of the (m + 1) -th to n-th accent phrases.

[0013]

According to the speech synthesis method of the present invention, when estimating the difference of the pitch frequency of the center of gravity between each mora of an arbitrary character string at least according to the parameter information of the accent type or the parameter information of the number of mora, With respect to the first mora of the character string, pitch difference estimation processing is performed based on the difference between the pitch frequency of a fixed value and the pitch frequency of the center of gravity of the first mora.

Further, in the speech synthesis method of the present invention, a difference in pitch frequency of the center of gravity between each mora of an arbitrary character string is estimated corresponding to at least accent type parameter information or mora number parameter information. When the accent phrase including the leading mora of the character string is located at the beginning of the delimited range, the accent type of the preceding accent phrase of the accent phrase is set to the relief type or the flat type, and pitch difference estimation processing is performed. Do.

According to Sanseido, the second edition of the Meiji Japanese Accent Dictionary, Kazue Akinaga, a part of a single accent phrase that has a portion from a mora pronounced high to a mora pronounced low is called an undulating expression. Also, an accent phrase in which there is no portion from a mora that pronounces high to a mora that pronounces low is called a flat type.

Further, the position of the accent phrase in the sentence is also used as the parameter information, based on the pitch frequency difference between each mora of the sentence composed of m accent phrases, based on the pitch frequency of the centroid point between two adjacent mora. When estimating the difference between
In order to estimate the pitch frequency difference of a sentence composed of n (> m) accent phrases, the positions of the (m + 1) th to nth accent phrases in the sentence are set to m, and the (m + 1) th to nth accent phrases are set. Of the pitch frequency of the center of gravity point between each mora of the accent phrase of.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to FIGS.

FIG. 1 is a block diagram showing an embodiment of a speech synthesizing method according to the present invention. 1 is a character code symbol string input section for inputting a character string to be subjected to ruled speech synthesis, and 2 is a character string which is converted into a word. A morphological analysis unit that divides into units and determines part of speech information, 3 is a reading determination unit that determines the reading of the word, 4 is a word dictionary that stores the reading of the word, and 5 is an accent that is determined based on the reading of the word An accent deciding unit 6 stores an accent dictionary storing accents for each word, and a phrase deciding unit 7 determines a phrase of the character string. The phrase refers to an exhalation paragraph such as a sentence or a reading point, a reading point or a reading point, a reading point or a punctuation mark, a breath or a breath, or a pause or a pause.

Reference numeral 8 denotes a pitch pattern generation unit for generating the pitch pattern of the character string. Reference numeral 9 denotes a difference between the pitch frequency of the vowel centroid of the mora of the character string and the pitch frequency of the vowel centroid of the preceding mora with respect to the mora. , A pitch difference estimating unit for estimating based on a neural network or a correspondence table, 10 is a unit connecting unit for connecting speech units, 11 is a unit table storing speech units, 12 is a DA converter, 1
3 is a speaker.

FIG. 2 shows the result of morphological analysis of the input character string.

FIG. 3 shows the result of a reading decision made on the morphological analysis result of the input character string.

FIG. 4 shows an input character string in units of accent phrases. 41 is a second mora of a second accent phrase of the input character string, 42 is a first accent phrase of the input character string, and 43 is an input character string. This is the second accent phrase of the character string.

FIG. 5 shows a pitch pattern of an input character string. Here, in an accent phrase, an accent type in which an accent exists in the Nth mora is referred to as “N type”. Therefore, in the character string "Michi-no-kaze", the first accent phrase "michi-o" has no accent, so the accent type is 0, and the second accent phrase "michi-no" is accented on the third mora "ne". Exists, the accent type becomes type 3.

FIG. 6 is a correspondence table in which a plurality of short sentences of one phrase composed of two accent phrases are used, and a difference in pitch frequency of a vowel barycenter between two adjacent moras is associated with each of five parameter information. is there.

FIG. 7 schematically shows the pitch difference estimating unit 9 using a neural network, wherein 71 is an input layer consisting of 5 units and 1 layer, 72 is an intermediate layer consisting of 10 units and 3 layers, and 73 Is an output layer consisting of one unit and one layer. Non-linear processing using a sigmoid function is used for the intermediate layer 72.

FIG. 8 shows an example of a pitch pattern of learning data to be learned by the neural network of the pitch difference estimating section 9.

FIG. 9 shows that the neural network learns the pitch frequency difference between the vowel centroids between each mora of the character string of one phrase composed of two accent phrases, and then converts the character string of one phrase composed of three accent phrases. FIG. 3 shows a pitch pattern when input to the character code symbol string input unit 1 to synthesize a rule speech.

First, learning data composed of various character strings is input to the input layer 71 by using a plurality of short sentences of one phrase composed of two accent phrases in the pitch difference estimating unit 9 of FIG.
Let them learn.

More specifically, a character string of one phrase consisting of two accent phrases shown in FIG.
The case where the neural network learns the difference between the pitch frequencies of the vowel centroids of the seventh mora "ru" and the sixth mora "ku" immediately before will be described.

In the neural network used in the pitch difference estimating unit 9 of this embodiment, the input layer 71 has the following information: 1) the position in the sentence of the accent phrase including the target mora 2) the number of mora of the accent phrase including the target mora 3) the target Position of the mora in the accent phrase 4) Accent type of the accent phrase including the focused mora 5) Accent type of the preceding accent phrase of the accent phrase including the focused mora 5 pieces of parameter information are input.

When the learning data composed of the above-mentioned five pieces of parameter information is input to the input layer 71, the neural network receives from the output layer 73 the pitch frequency of the center of gravity of the vowel of the target mora and the mora of the preceding mora. It has been learned to output the difference from the pitch frequency of the vowel centroid.

Considering the character string of one phrase "Taifu ga kuru" consisting of two accent phrases, the seventh mora "Ru", which is the attention mora, has a mora number of 2 The second of the second accent phrase in which the accent type is 1
Therefore, the parameter information to the input layer 71 of the neural network includes the position “2” in the sentence of the accent phrase including the target mora, the number of mora of the accent phrase including the target mora “2”, and the The position of the accent phrase in the accent phrase is “2”, the accent type of the accent phrase including the target mora is “1”, and the accent type of the preceding accent phrase of the accent phrase including the target mora is “3”, and the parameter information is “2, 2, 2, 1, 3 ".

By the way, in the present method, since the pitch frequency difference of the vowel centroid point between the target mora and the preceding mora is estimated, the first mora of the first accent phrase (the first mora at the beginning of the sentence). A problem arises in estimating the difference between the pitch frequency of "ta" and the preceding mora, and how to handle the accent type of the preceding accent phrase when the accent phrase of the focused mora is the first accent phrase.

To estimate the difference between the pitch frequency of the vowel center of gravity of the first mora "ta" of the first accent phrase and the preceding mora preceding this mora, the present embodiment uses The first of the first accent phrase other than type 1
Assuming that the average pitch frequency of the mora and the accent type of the preceding accent phrase with respect to the first mora of the first accent phrase belong to the undulation formula, the pitch frequency of the vowel centroid with this value and the first mora of the first accent phrase To determine the difference.

This is because the average pitch frequency of the first mora of the first accent phrase other than the first mora in the learning data is used as the preceding mora of the first mora of the first accent phrase because the accent type is the first mora. The first of the first accent phrase other than
This is because the value of the average pitch frequency of the mora is empirically low, and the type 1 is adopted as the accent type of the preceding accent phrase for the first mora of the first accent phrase. This is because the value of the pitch frequency decreases in the rear part of the accent phrase.

As a result, the connection from the preceding mora of the first accent phrase to the first mora to the first mora of the first accent phrase has no uncomfortable feeling and can be regarded as a natural voice utterance.

Thus, when the target mora corresponds to the first mora of the first accent phrase, the “accent of the preceding accent phrase of the accent phrase including the target mora” input to the input layer 71 of the neural network of the pitch difference estimator 9 is input. "Type" is changed to type 1.

By the way, as the preceding mora to the first mora of the first accent phrase (the first mora at the beginning of the sentence), for example, the average pitch frequency of the first mora of the first accent phrase of only the type 1 accent in the learning data is adopted. It is also possible. In this case, since the value of the average pitch frequency tends to be empirically high, the value of the first accent phrase
By setting the accent type of the preceding accent phrase to the mora to be, for example, a flat plate type in which the pitch of the rear portion is small, the connection from the preceding mora to the first mora of the first accent phrase to the first mora of the first accent phrase is made. Will feel uncomfortable.

As described above, the character string "Taifuga comes"
When attention is paid to the first mora “ta”, the five pieces of parameter information input to the input layer 71 include the position “1” in the sentence of the accent phrase including the attention mora, and the number of mora of the accent phrase including the attention mora “ 5 ”, the position“ 1 ”in the accent phrase of the target mora, the accent type“ 3 ”of the accent phrase including the target mora, and the accent type“ 1 ”of the preceding accent phrase of the accent phrase including the target mora. It is represented by parameter information “1, 5, 1, 3, 1”. By inputting these five parameters “1, 5, 1, 3, 1” to the input layer 71 of the neural network,
The difference between the vowel center-of-gravity point pitch frequency of La "ta" and the preceding mora is estimated.

Therefore, in the above-described neural network, while the learning data of a plurality of character strings having different accent types and mora numbers represented by the above five parameters is input to the input layer 71, the error back propagation is performed. The strength of the connection between the units of the neural network until the output value of the neural network converges in the vicinity of the desired pitch frequency difference, that is, the value of the pitch difference obtained from the pitch pattern of natural speech ( Weighting) is sequentially changed, and learning is repeated.

Thus, the strength of the connection between the units of the neural network is determined.

In the following, an unlearned character string "Ask for a road" is converted into a rule speech by using a pitch difference estimator 9 composed of a neural network in which the final connection strength between the units is determined. The following describes the processing operation in the case of causing the processing.

The character string input from the character code symbol string input unit (1) is first divided by the morphological analysis unit 2 into word units, and the part of speech of each word is determined as shown in FIG. .

When the part of speech is determined by the morphological analysis unit 2,
The part-of-speech data is sent to the reading determination unit 3, and the reading of each word is determined by collation with the word dictionary 4 as shown in FIG. 3.

When the reading of the word is determined by the reading determining unit 3, the word data is sent to the accent determining unit 5, the accent of the word is determined by collation with the accent dictionary 6, and the accent connection is performed according to the rules. Then, an accent phrase is formed, and the accent for the accent phrase is determined. As a result, the character string “Ask Michio” is divided into two accent phrases, a first accent phrase “Michio” 42 and a second accent phrase “Teach”, as shown in FIG.

After the accent phrase and the accent are determined, the phrase determination unit 7 determines the phrase.
In the character string according to the present embodiment, one phrase is formed by the entire character string. For example, if the character string is “Go to Koban and ask for the way”, “Go to Koban” and “Michito” Ask ".

Next, the pitch difference estimation unit 9 estimates the pitch difference. In other words, in the case of the character string of the present embodiment, the fifth mora “Z”, which is the target mora, is the second mora of the second accent phrase 43 having the number of mora of 4 and the accent type of 3. −
La 41. Therefore, the five pieces of parameter information input to the input layer 71 include the position “2” in the sentence of the accent phrase including the target mora, the number of moras of the accent phrase including the target mora “4”, and the accent phrase of the target mora in the sentence. 5 parameter information "2,4,2,3" of the position "2", the accent type "3" of the accent phrase including the attention mora, and the accent type "0" of the preceding accent phrase of the accent phrase including the attention mora , 0 ". By inputting these five parameters “2,4,2,3,0” to the input layer 71 of the neural network, the center of gravity of each vowel center of the target mora “zu” and the preceding mora “ta” is obtained. The difference between the pitch frequencies is a natural logarithm and is estimated to be “+0.147”.

In this manner, for each of the “Miss the Way” models, the difference between the pitch frequency of the vowel barycenter of the target model and the pitch frequency of the vowel barycenter of the preceding model is the natural value. Logarithmically, "-0.061, 0.396,-
0.224, −0.300, 0.147, −0.142, −0.320 ”.

Thereafter, the pitch pattern generator 8 calculates the vowel center point of each mora based on the preset pitch frequencies at the start and end of the voice section and the values estimated by the pitch difference estimator 9. Is estimated, and a point pitch pattern shown in FIG. 5 is generated.

When the point pitch pattern is generated, the segment connection unit 10 pre-stores speech segments such as CVC (consonant + vowel + consonant) (for example, Percall coefficients or LSP coefficients). From the table 11, speech segments necessary for speech utterance of the input character string are selected, the respective segments are connected, and speech data as a digital signal is created. The audio data is converted into an analog signal by the DA converter 12 and is output from the speaker 13 as a regular synthesized voice.

In the above, learning of the neural network constituting the pitch difference estimating unit 9 is performed using a character string composed of two accent phrases, and a character string to be actually subjected to rule speech synthesis is also assumed to be composed of two accent phrases. However, in the following, a case will be described where the learning of the neural network is the same as described above with two accent phrases and a character string composed of three accent phrases is synthesized in a regular speech.

FIG. 9 shows an estimation and a pitch pattern of a natural voice when a character string composed of three accent phrases, for example, “a good friend gathers” is input to the character code symbol string input unit 1. Shown in

The third accent phrase “Atsumaru” is input to the neural network of the pitch difference estimating unit 9 (5).
Of the parameter information, the “position in the sentence of the accent phrase including the attention mora” is actually “3”,
In this embodiment, it is “2”.

This is because when observing the pitch frequency as a human voice utterance, the shape of the pitch frequency has an upwardly convex mountain shape, and the “position in the sentence of the accent phrase including the focused mora” is changed to “ It has been confirmed by the inventor's experiment that the setting of "2" is closer to natural voice utterance. Therefore, "the position of the accent phrase including the focused mora in the sentence" is set to "2".

As described above, in order to estimate the difference in pitch frequency between each mora of a sentence composed of the number of unlearned accent phrases, within one phrase, a sentence having m accent phrases is m. Using a neural network that has learned the difference between the pitch frequencies of the vowel centroids between each mora, n
When estimating the pitch frequency difference of the vowel barycenter between each mora of a sentence composed of (> m) accent phrases, (m +
1) For each mora of the nth to nth accent phrases, “position in sentence of accent phrase including attention mora”
Is estimated as m, it is possible to realize a rule-synthesized speech that is almost natural.

Further, as shown in the figure, although the pitch frequency estimation pattern of the third accent phrase is different from the pitch frequency of natural speech, in each mora of the (m + 1) th to nth accent phrases, " By estimating m as the position of the accent phrase including the focused mora in the sentence as m, and then multiplying the estimated value by a coefficient to make correction, the sound becomes closer to natural speech.

In the above embodiment, the pitch difference estimating unit 9 composed of a neural network changes the strength of the connection between the units while learning using a plurality of short sentences. Using a neural network that stores the determined strength of connection between each unit, the pitch frequency difference between vowel centroids between each mora was estimated. It goes without saying that the five parameter information and the difference between the pitch frequencies of the vowel centroids among the mora may be stored in a correspondence table as shown in FIG.

In this case, the character string "Michio
Paying attention to the fifth mora “zu” of “question”, the five pieces of parameter information become “2, 4, 2, 3, 0” 60, and the vowel centroid of the attention mora “zu” and the preceding mora “ta” The difference between the pitch frequencies of the points is a natural logarithm, and is “+0.147” 61 as in the case of the neural network.

In the present embodiment, the input layer 71 of the neural network has one unit of 5 units, and the intermediate layer 72 has three layers of 10 units. However, the number of units and layers are not limited to these.

In the above embodiment, the pitch frequency difference is estimated by focusing on the vowel center of gravity between two adjacent moras. However, the present invention is not limited to this, and the pitch frequency difference between the center of gravity of each mora is estimated. It is needless to say that may be estimated.

Further, in the present embodiment, as learning data, 2
Learning neural networks and creating correspondence tables using short sentences composed of accent phrases, but not limited to this, learning neural networks and creating correspondence tables using sentences with three or more accent phrases. Needless to say, it can be done.

In the above embodiment, the neural network learned by inputting five parameters, or 5
A pitch pattern of a character string composed of one phrase was generated using a correspondence table composed of three parameters.
Language information parameters in addition to the two parameters,
For example, whether or not the focused mora is unvoiced, whether or not the focused mora is a repellent sound, whether or not the focused mora is a relentless sound, whether or not the focused mora is accompanied by a voiced consonant, or a consonant of the focused mora is a fricative sound Whether or not the consonant of the target mora is a semi-vowel, whether or not the consonant of the target mora is a nasal sound, whether or not the consonant of the target mora is an affricate, the consonant of the target mora is a plosive Neural network trained by adopting whether or not there is, whether or not the part of speech of the word containing the noticed mora, or whether the accent phrase containing the noticed mora is emphasized, or a correspondence table May be used.

A pitch pattern of a character string of a plurality of phrases is generated by a neural network or a correspondence table learned using five parameters, parameters relating to the above-mentioned linguistic information, parameters such as a phrase position or the number of phrases. It is also possible.

[0064]

As described above, according to the present invention, when estimating the difference of the pitch frequency of the center of gravity between the mora of the character string, when the mora of interest is located at the head of the divided range, Assuming a constant pitch frequency as the leading mora of the target mora, estimating the difference between the pitch frequency of the center of gravity and the mora, and at this time, the accent phrase including the target mora is placed at the beginning of the delimited range. If it is located, the rule-type synthesized speech can be made closer to a natural speech by estimating the difference in pitch frequency by making the accent type of the preceding accent phrase of the accent phrase an undulating or flat type.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of rule speech synthesis using the present invention.

FIG. 2 is a diagram showing a result of morphological analysis of an input character string;

FIG. 3 is a diagram showing a result of determining reading of a morphological analysis result of an input character string;

FIG. 4 is a diagram showing an input character string in units of accent phrases.

FIG. 5 is a diagram showing a pitch pattern of an input character string.

FIG. 6 is a corresponding chart used in the pitch difference estimating unit 9;

FIG. 7 is a configuration diagram of a neural network used in a pitch difference estimating unit 9;

FIG. 8 shows an example of a pitch pattern of learning data to be learned by a neural network of the pitch difference estimating unit 9;

FIG. 9 shows a pitch pattern when a sentence of three accent phrases is synthesized into a regular speech using a neural network trained on learning data of two accent phrases.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character code symbol string input part 2 Morphological analysis part 3 Reading determination part 5 Accent determination part 6 Accent dictionary 7 Phrase determination part 8 Pitch pattern generation part 9 Pitch difference estimation part 10 Element connection part 11 Element table 71 Input of neural network Layer 72 Intermediate layer of neural network 73 Output layer of neural network

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-59-81697 (JP, A) JP-A-59-192293 (JP, A) JP-A-61-70597 (JP, A) JP-A-62 262100 (JP, A) JP-A-63-85797 (JP, A) JP-A-1-238697 (JP, A) JP-A-4-46396 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 13/06-13/08 G10L 21/04 G06F 15 / 18,17 / 27

Claims (13)

(57) [Claims] 1. A pitch difference estimating process for estimating a pitch frequency difference of a center of gravity between two adjacent moras corresponding to at least accent type parameter information or mora number parameter information; On the basis of the,
A pitch pattern generation process for generating a pitch pattern of an arbitrary character string; and a speech unit necessary for speech utterance of the character string from a speech unit memory in which a speech unit including speech parameters necessary for speech synthesis is stored. Read out and connect,
A voice unit connection process for generating voice data, wherein when a certain mora is located at the beginning of the divided range, the pitch difference estimating process uses a constant pitch frequency. And estimating a difference between the pitch frequency of the center of gravity of the mora and the pitch frequency of the mora. 2. The pitch difference estimating process includes: when at least accent type parameter information or mora number parameter information is input to the input layer, the pitch difference estimating process is performed on the adjacent layer from the output layer.
Based on a neural network that has been trained to output the difference between the pitch frequencies of the centroid points between the mora,
This neural network is characterized in that when a certain mora is located at the beginning of a divided range, it has been learned to output a difference between a pitch frequency of a fixed value and a pitch frequency of the center of gravity of the mora. The speech synthesis method according to claim 1. 3. The pitch difference estimating process uses a correspondence table that stores a difference in pitch frequency of a center of gravity between two adjacent moras corresponding to at least accent type parameter information or mora number parameter information. Then, the difference between the pitch frequencies is estimated, and the correspondence table stores the difference between the pitch frequency of a fixed value and the pitch frequency of the center of gravity of the mora when a certain mora is located at the beginning of the divided range. 2. The speech synthesis method according to claim 1, wherein: 4. At least a difference in pitch frequency of the center of gravity between two adjacent moras corresponding to the parameter information of the accent type of the preceding accent phrase of the accent phrase containing a certain mora or the parameter information of the number of mora. Based on the pitch difference estimation process to be estimated and the difference between the pitch frequencies,
A pitch pattern generation process for generating a pitch pattern of an arbitrary character string; and a speech unit necessary for speech utterance of the character string from a speech unit memory in which a speech unit including speech parameters necessary for speech synthesis is stored. Read out and connect,
In a speech synthesis method comprising speech unit connection processing for generating speech data, when an accent phrase including a certain mora is located at the beginning of a delimited range, the accent of the preceding accent phrase of the accent phrase is A voice synthesizing method characterized in that the mold is of an undulating or flat type. 5. The pitch difference estimating process is characterized in that, when at least accent type parameter information of a leading accent phrase of an accent phrase including a certain mora or parameter information of the number of mora is input to the input layer, the pitch difference estimation processing is performed from the output layer. Do 2
Based on a neural network that has been trained to output the difference between the pitch frequencies of the centroid points between the mora,
When the accent phrase including a certain mora is located at the beginning of the divided range, the accent type of the preceding accent phrase of the accent phrase is set to the undulating type or the flat type, and the difference of the pitch frequency is added to the neural network. 5. A learning is performed so as to output.
Described speech synthesis method. 6. The pitch difference estimating process includes:
Using a correspondence table that stores the difference between the pitch frequencies of the centroid points of two adjacent mora corresponding to the parameter information of the accent type of the preceding accent phrase of the accent phrase including a certain mora or the parameter information of the number of mora. , The pitch frequency difference of the center of gravity between two adjacent moras is estimated, and if the accent phrase including a certain mora is located at the beginning of the delimited range, the leading accent of the accent phrase is 5. The speech synthesis method according to claim 4, wherein the accent type of the phrase is set as a relief type or a flat type, and the difference between the pitch frequencies is stored in the correspondence table. 7. The speech synthesis method according to claim 1, wherein the divided range is a sentence. 8. The speech synthesis method according to claim 1, wherein the divided range is a phrase. 9. The speech synthesis method according to claim 1, wherein the divided range is an accent phrase. 10. The speech synthesis method according to claim 1, wherein the divided range is a word. 11. According to at least a position of an accent phrase in a sentence, parameter information of an accent type, or parameter information of a number of mora, a pitch frequency difference between mora of a sentence composed of m accent phrases is determined. A pitch difference estimating process for estimating a pitch frequency difference between centroid points between two adjacent moras based on the pitch frequency generating process for generating a pitch pattern of an arbitrary character string based on the pitch frequency difference; Speech unit connection processing for reading and connecting the speech units required for speech utterance of the character string from a speech unit memory in which speech units composed of speech parameters necessary for speech synthesis are stored, and generating speech data When estimating the pitch frequency difference of a sentence composed of n (> m) accent phrases, the (m + 1) th to the (m + 1) th The position in the context of th accent phrase as m, (m + 1)
A speech synthesis method characterized by estimating a difference between pitch frequencies of centroid points between mora of the nth to nth accent phrases. 12. The pitch difference estimating process includes, for each mora of a sentence composed of m accent phrases, a parameter of at least a position of an accent phrase in the sentence, an accent type, or a number of mora to an input layer of a neural network. 12. The speech synthesis method according to claim 11, wherein when information is input, the speech synthesis method is performed based on a neural network learned to output a pitch frequency difference between centroid points between two adjacent moras from an output layer. 13. The pitch difference estimating process includes, for each mora of a sentence composed of m accent phrases, at least:
12. The speech synthesis method according to claim 11, wherein the speech synthesis method is performed based on a correspondence table storing parameter information of a position of an accent phrase in a sentence, an accent type, or a number of mora.