JP3485586B2 - Voice synthesis method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rule-based speech synthesis method, and more particularly, to a method for generating a pitch pattern of synthesized speech which greatly affects the naturalness of the phoneme and accent of the speech.

[0002]

2. Description of the Related Art In recent years, research on speech synthesis by rules has been actively carried out as an important technique for human interface. Speech synthesis by rules determines part-of-speech information by morphological analysis from character strings and the like, and after determining the reading of a word by matching with a word dictionary, obtains the accent type, accent combination, and phrase of the word according to this reading, A pitch pattern is determined from these pieces of information, and at the same time, a speech unit (for example, a Percoll coefficient or LS) corresponding to the reading of a word is determined.
The audio data is generated by connecting (P coefficient).
That is, the voice data is a Percoll coefficient sequence, a pitch pattern and amplifier information corresponding thereto. Among them, the pitch pattern is positioned as having a great influence on whether the naturalness of the synthesized voice is good or bad.

In the conventional pitch pattern generation, after determining a phrase component and an accent component from the structure of a sentence,
A method using 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 is known. (IEICE Transactions Vol.
J63-D No.9 pp.715-722, 1980.9) Also known is a method of estimating the value of each pitch frequency of the leading mora, the mora having a peak pitch frequency, and the ending mora for each phrase by a neural network. There is. (Voice study group material SP89-111, 1990.1) All of these conventional pitch pattern generation methods are methods for uniquely estimating the absolute value of the pitch frequency for each mora, and the connection with the mora before and after (pitch The amount of frequency change) is not taken into consideration. Therefore,
These methods have a problem that the amount of change in pitch pattern between mora is not stable.

[0004]

As described above, although the variation amount is more important than the absolute value of the pitch frequency as a factor that influences the naturalness of the synthesized speech, the conventional method is used. Neither estimate the variation in pitch frequency for each mora.

The present invention has been made to solve such a problem, and in speech synthesis by rules,
In order to improve the naturalness of synthetic speech, a pitch pattern is generated in consideration of the amount of change in pitch frequency that greatly affects the naturalness of synthetic speech.

[0006]

According to the pitch pattern generation method of the present invention, for each mora of a character string, the pitch frequency of the center of gravity of the mora and the pitch frequency of the center of gravity of the preceding mora with respect to the mora are calculated. A pitch difference estimation process for estimating a difference, and a pitch pattern generation process for forming a pitch pattern from the pitch difference obtained by the pitch difference estimation process, wherein the pitch pattern generation process is at least the Accent phrase with mora
Mora position of the mora in question , divided phrases
Accent phrase position including the mora in, the motor
The number of mora in the accent phrase containing the
Accent type of quint phrase, Accen containing the mora
From the correspondence between the information about the accent type of the preceding phrase and the difference in pitch frequency between the concatenated moras centroids, the pitch frequency of the centroid of each mora and the pitch frequency of the centroid of the preceding mora Find the pitch difference between
A feature is that a pitch pattern is generated based on the obtained pitch difference.

Further, the voice synthesis method of the present invention is a pitch pattern generation process for generating a pitch for each mora of an input arbitrary character string, and a voice unit for storing voice units including voice parameters necessary for voice synthesis. One memory,
A voice unit connection process for connecting the voice unit required to the character string to generate voice data, wherein the pitch pattern generation process is at least the mora.
The mora position of the mora in the accent phrase containing
Accent phrase position including the mora in the divided phrases, mora number of accent phrases including the mora,
Accent type of the accent phrase including the mora, the motor
Accenture of preceding accent phrase including accent
The pitch difference between the center frequency of the corresponding mora and the center frequency of the preceding mora is calculated based on the correspondence between the pitch type information and the pitch frequency difference between the adjacent mora center points. A pitch pattern is generated based on the obtained pitch difference.

Further, the voice synthesis method of the present invention is a pitch pattern generation process for generating a pitch for each mora of an input arbitrary character string, and a voice unit for storing a voice unit including voice parameters necessary for voice synthesis. A speech unit connection process for connecting one-sided memory and the voice unit necessary for the character string to generate voice data, wherein the pitch pattern generation process includes at least an accent including the mora. mora position of the mora definitive in clause division
Been accent phrase position including the mora in phrases, number of moras accent phrase including the mora, the accent type of the accent phrase including the mora, and information about the accent type of the preceding accent phrase accent phrase including the mora, The pitch difference between the pitch frequency of the center point of gravity of the relevant mora and the pitch frequency of the center point of gravity of the preceding mora is calculated based on the correspondence relationship with the difference in pitch frequency between the concentric points of the moras. It is characterized in that a pitch pattern is generated based on this.

[0009]

In the speech synthesis method of the present invention, first, for each mora of an arbitrary character string, the difference between the pitch frequency of the center of gravity of the mora and the pitch frequency of the center of gravity of the preceding mora with respect to the mora is estimated. Such a difference estimation is at least based on the mood in the accent phrase containing the mora.
La Mora position, the mode in divided phrases
Accent phrase position including La, Accen including the mora
Number of mora of the phrase, accent type of accent phrase containing the mora, preceding access of accent phrase containing the mora
It is obtained from the correspondence between the information about the accent type of the phrase and the difference in the pitch frequency between the concatenated mora centroids. For example, the correspondence between the above information and the difference between the pitch frequencies. Is recorded, or when each of the above information is input to the input layer, the output layer is performed using a neural network or the like that has been learned to output the pitch difference between moras.

Next, by the above estimation, it is possible to generate the generation of the pitch pattern using the pitch difference for each concatenated mora.

[0011]

First, a rule synthesizing apparatus using the speech synthesizing method of the present invention will be described. [Embodiment 1] FIG. 1 is a block diagram showing an embodiment of a rule synthesizing apparatus using the speech synthesizing method of the present invention. In FIG. 1, 1 is a character code symbol string input unit for inputting a character string for regular speech synthesis, 2 is a morphological analysis unit that divides the character string into word units and determines part-of-speech information, and 3 is a reading of the word. A phonetic deciding unit for deciding a word reading, a word dictionary storing a word reading, a reference numeral 5 for an accent deciding unit for deciding an accent based on the word reading, a accent dictionary 6 for storing an accent for each word, Is a phrase determination unit that determines the phrase of the character string. The phrase refers to an exhalation paragraph such as a sentence head or a reading point, a reading point or a reading point, a reading point or a phrase, a breath or breath, or a pause or a pose.

Reference numeral 8 is a pitch pattern generator for generating the pitch pattern of the character string, and reference numeral 9 is a pitch difference estimation for estimating a difference between the pitch frequency of the center of gravity of the mora and the pitch frequency of the center of gravity of the preceding mora with respect to the mora. The unit, 10 is a unit connection unit for connecting voice units, 11 is a voice unit table storing voice units, 12 is a DA conversion unit, and 13 is a speaker.

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

FIG. 3 shows the result of 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. In FIG. 4, 41 is the fifth mora of the input character string, 42 is the first accent phrase of the input character string, 4
3 is the second accent phrase of the input character string.

FIG. 5 shows the pitch pattern of the input character string.

FIG. 6 shows a pitch difference estimating unit (9) using a correspondence table. In FIG. 6, 60 is an example of the mora position for the accent phrase including the mora, 61 is an example of the position of the accent phrase including the mora, 62 is an example of the number of mora of the accent phrase including the mora, and 63 is the mora. An example of the accent type of the accent phrase, 64 is an example of the accent type of the preceding accent phrase of the accent phrase including the mora, and 65 is an example of the pitch frequency difference between the mora and the preceding mora.

The processing operation of this embodiment will be described below with reference to a character string "ask the way" of one phrase consisting of two accent phrases.

The character string input from the character code symbol string input unit (1) is divided into word units by the morpheme analysis unit (2), and the part of speech of each word is determined. FIG. 2 is a result of dividing the character string “Ask the road” of this embodiment into word units and assigning a part of speech to each word.

When the part-of-speech is determined, it is sent to the reading determination unit (3) and the reading of each word is determined by collating with the word dictionary (4). In FIG. 3, the character string “Ask the road” in this embodiment is used.
The result of reading decision is shown for.

When the reading of the word is determined, it 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 combination is performed by the rule to form the accent phrase. As well as
The accent for the accent phrase is determined. As a result, the character string "Ask for the road" in this embodiment is, as shown in FIG. 4, the first accent phrase "Michio" (42).
And the second accent phrase “Tsuneru” (43).

After the accent is determined, the phrase is determined by the phrase determining section (7). In the character string of the present embodiment, one phrase is formed by the entire character string, but for example, "I went to the koban and asked the road".
If it is a character string such as, it is divided into two phrases, "I'm going to ask" and "I asked Michi."

Next, the pitch difference is estimated. In the present embodiment, the pitch frequency of the center of gravity of the fourth moor "ta" and the fifth moor "zu" of the character string "Ask for Michi". A case will be described in which the difference is estimated using the correspondence table of FIG.

Here, the correspondence table is: 1) the mora in the accent phrase including the mora
Mora position 2) accent phrase position 3 containing the mora in divided phrases) accent type 5 of the mora number 4 accent phrase including Mora) accent phrase containing the mora) prior accent accent phrase including the mora The pitch difference between the mora and the preceding mora is described in natural logarithm in a form corresponding to the five accent type parameters of the phrase. Such a correspondence table is created, for example, by using a large number of sentences and recording the pitch difference between the mora of the sentence under the condition that the accent type and the number of mora have various values.

In the case of the character string of this embodiment, the fifth mora "zu", which is the mora, is the second mora (43) of the second accent phrase (43) in which the number of mora is 4 and the accent type is 3. Four
1). In the correspondence table, this is the mora position “2” (60) for the accent phrase containing the mora,
The accent phrase position “2” (61) including the mora, and the mora number “4” (6) of the accent phrase including the mora.
2), the accent type “3” (63) of the accent phrase including the mora, the accent type “0” (64) of the preceding accent phrase (42) of the accent phrase including the mora,
It is represented by five parameters “2, 2, 4, 3, 0”. Therefore, from the correspondence table, it is estimated that the difference between the center frequency pitch frequencies of the mora "zu" and the preceding mora "ta" is "+0.147" (65) in natural logarithm.

By the way, in this method, since the difference between the pitch frequencies of the vowel centroids of the target mora and the preceding mora preceding it is estimated, the first mora of the first accent phrase (the first mora at the beginning of the sentence). And how to deal with the estimation of the preceding mora that precedes this.

Therefore, in the case of estimating the difference in pitch frequency of the vowel center of gravity points between the first mora of the first accent phrase and the preceding mora preceding this mora, the first accent phrase of the first accent phrase is used in this embodiment. The accent type of the preceding accent phrase with respect to the mora is set as type 1, and the difference between this value and the pitch frequency of the vowel center point of gravity with respect to the first mora of the first accent phrase is determined.

This is because the type 1 is adopted as the accent type of the preceding accent phrase with respect to the first mora of the first accent phrase because the value of the pitch frequency falls in the rear part of the accent phrase of the type 1 type. And as a result, from the preceding mora for the first mora of the first accent phrase,
The connection of the first accent phrase to the first mora has no discomfort and can be regarded as a natural voice utterance.

In this way, for each mora of "Ask for the road", the difference between the center-of-mass pitch frequency of the mora and the center-of-mass pitch frequency of the preceding mora is the natural logarithm,
From the first mora, "-0.061, 0.396, -0.224,-"
0.300, 0.147, −0.142, −0.320 ”.

In the pitch pattern generator (8), the center of gravity of each mora is set based on the preset pitch frequencies of the start and end of the voice section and the respective values estimated by the pitch difference estimator (9). Estimating the pitch frequency at, the point pitch pattern as shown in FIG. 5 is generated.

When the pitch pattern is generated, the speech unit (for example, Percoll coefficient or LSP) such as CVC (consonant + vowel + consonant) is generated in the segment connecting portion (10).
Element table (11) in which coefficients are stored in advance
Then, the voice unit necessary for the sentence is selected, each unit is connected, and voice data which is a digital signal is created. The voice data is converted into an analog signal by the DA converter (12) and output as a synthesized voice from the speaker (13).

In the above-described embodiment, the pitch pattern of the character string consisting of one phrase is generated based on the correspondence table composed of the five parameters shown in FIG.
In place of or in addition to some of these five parameters, parameters relating to language information, for example, whether or not the attention mora is unvoiced, whether or not the attention mora is accompanied by unvoiced consonants, , Whether or not the focused mora is a jumble, whether or not the focused mora is accompanied by voiced consonants, and whether or not the consonant of the focused mora is a fricative,
Whether the consonant of the target mora is a half vowel, whether the consonant of the target mora is a nasal, whether the consonant of the target mora is an affricate, whether the consonant of the target mora is a plosive ,
The correspondence table may be created by adopting what the part of speech of the word including the attention mora is, whether the accent phrase including the attention mora is emphasized, or the like.

In the above embodiment, the pitch pattern of the character string consisting of one phrase is generated. However, the five parameters of the correspondence table of FIG. 6, the parameters relating to the language information, the phrase position, or the number of phrases are used. It is also possible to generate a pitch pattern of character strings of a plurality of phrases by a correspondence table using, for example. [Embodiment 2] Next, an embodiment in which a neural network is used in the pitch difference estimation unit (9) of Embodiment 1 will be described.

FIG. 7 shows a pitch difference estimating section (9) using a neural network. In FIG. 7, 71 is an input layer,
72 is an intermediate layer and 73 is an output layer.

Since the processes other than the pitch difference estimating unit (9) are the same as those in the processing operation 1, only the process in the pitch difference estimating unit will be described below.

In this embodiment, the difference between the pitch frequency of the center of gravity of the fourth moor "ta" and the fifth moor "zu" of the character string "ask for the road" of this embodiment expressed by reading is calculated. The case of estimation by a neural network will be described.

In the neural network used in the pitch difference estimation unit (9), the input layer (71) has: 1) the mora in the accent phrase including the mora.
Mora position 2) accent phrase position 3 containing the mora in divided phrases) accent type 5 of the mora number 4 accent phrase including Mora) accent phrase containing the mora) prior accent accent phrase including the mora Enter the five parameters of the phrase accent type.

In this neural network, it is assumed that the output layer (73) is learned so as to output the difference between the pitch frequency of the center of gravity of the mora and the pitch frequency of the center of gravity of the preceding mora. .

Here, since the fifth mora "zu" which is the mora is the second mora (41) of the second accent phrase (43) having the number of mora of 4 and the accent type of type 3, Input parameters to the input layer of the neural network are
The mora position “2” for the accent phrase including the mora, the accent phrase position “2” including the mora, the mora number “4” for the accent phrase including the mora, the accent type “3” for the accent phrase including the mora, The accent type “0” of the preceding accent phrase (42) of the accent phrase including the mora is “2, 2, 4, 3, 0”.

The neural network consists of the input layer (7
When 5 pieces of information "2, 2, 4, 3, 0" are input to 1), it is learned to output a certain value to the output layer (73), and is determined by this learning. Each unit of the neural network is weighted according to the existing coefficient. Thereby, from the output layer (73), for example, in natural logarithm, the difference "+0.147" between the center frequency pitch frequencies of the moor "zu" and the preceding mora "ta".
Is estimated and output.

Also in this embodiment, when the difference in the barycentric point pitch frequency between the first mora of the first accent phrase and the preceding mora is estimated by using a neural network, as in the first embodiment, for example, The accent type of the preceding accent phrase including the mora is learned as "1" (information that can be regarded as other than 0 type), and the preceding accent of the accent phrase including the mora among the input parameters to the input layer is learned. The phrase accent type is estimated as "1" (input information used during learning).

In this way, for each mora of "Ask for the road", the difference between the center frequency of the mora and the center frequency of the preceding mora is the natural logarithm,
From the first mora, “-0.061, 0.396, −0.224, −
0.300, 0.147, −0.142, −0.320 ”.

In the above second embodiment, the input layer of the neural network is 5 units and the intermediate layer is 1 layer, but the number of each unit and the number of layers are not limited to this.

Further, in the above-mentioned second embodiment, the pitch pattern of the character string consisting of one phrase is generated by using the neural net learned by inputting the five parameters. Or a parameter related to linguistic information in addition to these five parameters, for example, whether the attention mora is unvoiced, whether the attention mora is accompanied by unvoiced consonants, and whether the attention mora is sound-repellent. Or whether or not the attention mora is a roar,
Whether the attention mora is accompanied by voiced consonants, whether the attention mora's consonants are fricatives, whether the attention mora's consonants are half vowels, whether the attention mora's consonants are nasal sounds, Whether the consonant is an affricate, whether the consonant of the attention mora is a plosive, what the part of speech of the word containing the attention mora is, or whether the accent phrase containing the attention mora is emphasized It is also possible to use a neural network that has been trained by adopting ,, or the like.

Further, in the above-mentioned second embodiment, the pitch pattern of the character string consisting of one phrase is generated.
It is also possible to generate a pitch pattern of a character string of a plurality of phrases by a neural net learned using the above-mentioned parameters relating to language information, phrase positions, or the number of phrases.

[0046]

As described above, according to the present invention, for each mora, by estimating the difference between the pitch frequency of the center of gravity of the mora and the pitch frequency of the center of gravity of the preceding mora of the mora. , It is possible to improve the naturalness of synthesized speech by generating a pitch pattern from fluctuations in pitch frequency.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of rule 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 diagram showing a correspondence table used in a pitch difference estimation unit (9).

FIG. 7 is a block diagram of a neural network used in a pitch difference estimation unit (9).

[Explanation of symbols]

1 character code symbol string input section 2 Morphological analyzer 3 reading decision section 4 word dictionary 5 Accent determination part 6 accent dictionary 7 Phrase decision section 8 Pitch pattern generator 9 Pitch difference estimation unit 10 Element connection part 11 piece table 12 DA converter 13 speakers 41 5th mora of input character string "zu" 42 First accent phrase of input string 43 Second accent phrase of input string 71 Neural network input layer 72 Middle layer of neural network 73 Output layer of neural network

Front page continuation (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 13/06-13/08 JISST file (JOIS)

Claims (2)

(57) [Claims] 1. A pitch difference estimation process for estimating a difference between a pitch frequency of a center of gravity of the mora and a pitch frequency of a center of gravity of a preceding mora with respect to the mora, and the pitch difference estimation. in the pitch pattern generation method consisting of the pitch pattern generation process for forming a pitch pattern from the pitch difference determined by the processing, the pitch pattern generation process, at least, the motor
The mora position of the mora in the accent phrase including the la, the accent phrase position including the mora in the divided phrase, the number of mora of the accent phrase including the mora, the accent type of the accent phrase including the mora , This
The action of the preceding accent phrase of the accent phrase containing the mora
The pitch difference between the pitch frequency of the center of gravity of each mora and the pitch frequency of the center of gravity of the preceding mora is found and obtained from the correspondence relationship between the information about the cent type and the difference in the pitch frequency between the connecting points of the mora. A pitch pattern generation method characterized by generating a pitch pattern based on the obtained pitch difference. 2. A pitch pattern generation process for generating a pitch for each mora of an input arbitrary character string, a voice unit memory for storing a voice unit consisting of voice parameters necessary for voice synthesis, and the character string. A speech synthesis method comprising a speech segment connection process for connecting the required speech segments to generate speech data, wherein the pitch pattern generation process is at least a mora position of the mora in an accent phrase including the mora. Information about the position of the accent phrase including the mora in the divided phrase, the number of mora of the accent phrase including the mora, the accent type of the accent phrase including the mora, and the accent type of the preceding accent phrase of the accent phrase including the mora. Based on the correspondence between the pitch frequency difference between the conjoined mora centroids, Speech synthesis method characterized by seeking the pitch difference between the pitch frequency of the center of gravity of the mora and the preceding pitch frequency of the center of gravity of mora, to generate a pitch pattern based on the determined pitch difference.