JP3310226B2 - Voice synthesis method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis method and apparatus, and more particularly, to a speech synthesis method and apparatus for converting text into speech.

[0002]

2. Description of the Related Art Conventionally, a speech synthesis method for synthesizing speech by connecting speech segments, as described in JP-A- 10-171484, constructs a speech segment database irrespective of a pitch pattern to be controlled. Was.

[0003]

However, in the conventional speech synthesis method, there is no guarantee that a speech fragment of a pitch pattern similar to the pitch pattern of the control target exists in the speech fragment database. When the pitch pattern that the voice piece originally has is significantly different,
There has been a problem that sound quality deterioration due to the pitch change is conspicuous, resulting in sound quality deterioration of synthesized speech.

[0004] The present invention provides a speech synthesis method for preventing such sound quality deterioration due to pitch change, and also provides a speech synthesis apparatus in which the ambiguity of the speech piece database is removed from the viewpoint of sound quality deterioration due to pitch change. The purpose is to:

[0005]

In order to solve the above-mentioned problems, according to the present invention, a pitch pattern to be controlled is determined by a template, and this template is divided into sections of a synthesis unit, and the pitch of each section is calculated. The height and shape are determined in advance and classified into categories. Then, configure the speech piece database as audio piece of the above categories the speech piece database for each phoneme chain one always exists, in synthesizing a sentence is a sentence end
Modify the template to lower the pitch as
Correct and combine these modified templates
Use as a template for the whole chapter, lower the template
The level to be corrected is classified into several levels, and
Which category does the pitch and shape of the voice piece belong to?
Is determined in advance, and a speech piece is selected based on the category classification . As described above, since it is possible to guarantee that a voice pattern having a pitch pattern similar to the pitch pattern to be controlled exists in the voice pattern database, it is possible to suppress the sound quality deterioration due to the pitch change and improve the quality of the synthesized voice. Yes, and it can handle all sentences, such as long sentences, with a finite template.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS According to the first aspect of the present invention, a speech unit of a synthesis unit such as CV, CV / VC, VCV, or CV / VCV is selected from a database, and the pitch of the selected speech unit is selected. Is modified according to a template that is an approximate pattern of the pitch, and in the speech synthesis method of the waveform superposition method in which these modified speech pieces are connected to synthesize speech, the template that is the approximate pattern of the pitch is divided into sections of synthesis units. , each segment is categorized by the height and shape of the pitch constitutes a speech piece database as phoneme chain to cover all categories for each same audio piece, in synthesizing a sentence will endnotes As
Modify the template to lower the pitch
Combining the modified templates in
Use as a plate and adjust the template low
Audio pitches for each level.
Beforehand which category the height and shape belong to
This is a voice synthesis method characterized by selecting voice segments according to this category classification , and it is possible to limit the maximum value of pitch change, so that the lowest level of sound quality deterioration due to pitch change can be set. , Also finite
Template for all sentences such as long sentences
It has the effect that the Ru can.

[0007] The invention according to claim 2 attempts to synthesize.
Input means for inputting voice readings
From CV, CV / VC, VCV, CV / VCV
Phonological chain determining means for determining the phonological chain of the speech unit of the unit
And the pitch and shape of the pitch of the speech piece for each phoneme chain
Database of voice segments, and pick up words and sentences
The pitch template, which is the approximate pattern of the switch, was recorded.
About pitch template database and pitch control
The pitch pattern according to the pitch template
Prosody determining means and the pitch template
Divided into sections, and the pitch height and shape in each section
The category that records the information that the category belongs to
According to the category selection table and the category selection table
A voice segment selecting means for selecting a voice segment, and a pitch of the voice segment.
Speech piece deformation means deforming according to a pitch template
And voice unit connection means for connecting the deformed voice unit, and synthesis
Means for outputting synthesized speech,
A pitch template, which is an outline pattern of the pitch of a sentence
Set to the length of a word or compound word to combine sentences
If you use these pitch templates as the end of the sentence
The pitch pitch
Combine rates to create pitch templates for writing
And a means for generating a pitch template for a sentence to be generated.
Lower the pitch in each pitch template
Multiple category selection tables for each level of classification
The voice synthesizer is characterized by
Voice segments with less sound quality degradation by pitch category
To select a voice segment,
Yes, the amount of processing required during synthesis can be reduced, and
Pitch template and category selection for text that is size
Since there is no selection table, pitch control of the speech synthesizer
This has the effect that the size of the application data can be reduced .

[0008]

[0009]

[0010]

[0011]

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG. For Embodiment 1 implementation of the present invention (Embodiment 1) it will be specifically described. To synthesize speech, first, speech segments of a target phoneme are collected, and after modifying the prosody such as pitch and rhythm of these speech segments, these are connected to synthesize speech. At this time, if there is a large difference between the pitch pattern originally possessed by the voice piece and the pitch pattern of the control target, sound quality degradation due to the pitch change occurs. In the conventional speech synthesis method, there was no guarantee that a speech segment close to the target pitch pattern existed in the speech segment database. It is.

The first embodiment solves the above problem, and FIGS. 1, 2 and 3 are conceptual diagrams.
FIGS. 1, 2 and 3 show that the unit of the speech piece is VCV (vowel-
(Consonant-vowel). FIG. 1 is an example of a pitch template to be a control target of a pitch pattern.
Reference numerals 101 to 106 indicate fundamental frequencies for each mora position.
To synthesize speech, the pitch is changed so that the vowel parts at both ends of the VCV match the pitch template, and these are connected to synthesize speech. At this time, in order to ensure a voice segment similar to the pitch pattern to be controlled in the voice segment database, first, a pitch shape pattern in the pitch template is classified. FIG. 2 shows an example of pattern classification of the pitch shape. In FIG. 2, the horizontal axis is VCV
The fundamental frequency of the preceding vowel part of the voice segment is shown, and the vertical axis is the fundamental frequency of the following vowel part. The pitch shape of the voice segment 111 in FIG. 1 is such that the leading vowel is 150 Hz and the following vowel is
Since the frequency is 300 Hz, it corresponds to the position 201 in FIG. Similarly, speech segment sections 112 to 115 in FIG. 1 correspond to 202 to 205 in FIG.
Corresponding to the position. This is performed for all pitch templates. Then, category classification of the pitch shape is performed as indicated by 211 to 214 in FIG. Then, as shown in FIG. 3, a table is created for each pitch template, which describes the correspondence between the voice segment and the category of the pitch shape.

Next, the speech segment database is constructed so that there is always one speech segment at the central position of the category determined as shown in FIG. 2 for each phoneme chain. Then, when synthesizing voice, a voice segment is selected according to the table of FIG. As a result, since a voice piece having a pitch shape close to the pitch pattern to be controlled always exists in the voice piece database, it is possible to limit deterioration in sound quality due to pitch change.

In the first embodiment, the pitch template is used as a pitch control method. However, a pitch pattern is generated by a model, and the pitch shape of the pitch pattern generated by the model is classified into categories. The same effect can be obtained by constructing a speech segment database so as to include speech segments of the category.

[0016] Next (Embodiment 2), the implementation of the second embodiment of the present invention will be described. FIG. 4 is a conceptual diagram of a speech synthesis method according to the second embodiment. When changing the pitch of a voice segment, if the pitch is changed by about 0.3 octave, the sound quality does not deteriorate much in terms of hearing. Paying attention to this point, if the pitch shape categories are classified so that the range of the pitch change is 0.3 octave or less, it is possible to synthesize a sound in which the sound quality deterioration due to the pitch change is not perceivable. In FIG. 4, V1 represents the fundamental frequency of the preceding vowel part, and V2 represents the fundamental frequency of the succeeding vowel part.

In FIG. 4, V1 and V2 are 50
The memory is used every 0.3 octaves based on Hz. 401 to 413 indicated by bold lines indicate the category classification of the pitch shape. For example, for category 401, V1 is 173Hz
V2 is a category covering 261 Hz to 396 Hz. At this time, if there is a voice piece whose V1 is 212 Hz and V2 is 322 Hz, which is the center of this category, even if the pitch is changed within the range of this category, it is 0.
Since it is within three octaves, it is possible to prevent the sound quality deterioration due to the pitch change from being perceived. Therefore, a voice synthesis method in which sound quality deterioration due to a pitch change is not perceived by hearing can be realized by securing a voice shape having a pitch shape that becomes the center position of these categories in the voice piece database.

[0018] (Embodiment 3) Next, a third embodiment corresponding to claim 1 of the present invention will be described. FIG. 5 is a conceptual diagram of a speech synthesis method according to the third embodiment. In FIG. 5, 511 and 521
And 531 are pitch templates representing pitch patterns of words or compound words. When composing sentences,
These templates are combined to form a sentence pitch template. At this time, in the case of a sentence, the pitch tends to decrease as the sentence ends, so the pitch of the word or compound word is corrected to be 0.9 times or 0.8 times as low as the end of the sentence. These are combined to form a sentence pitch template.

In FIG. 5, pitch patterns 512 to 514
Represents a pitch template 511 obtained by multiplying the pitch template 511 by 0.9 to 0.7, and similarly, pitch templates 522 to 52
4 and 532 to 534 are also pitch templates 521 and
531 has been increased from 0.9 to 0.7 times. When synthesizing sentences, for example, pitch template 511,
By combining 522 and 533, it is possible to generate a pitch pattern in which the pitch decreases as the sentence ends.

On the other hand, pitch categories are classified into high (H), medium (N), and low (L) for the preceding vowel portion and the succeeding vowel portion, respectively, and a total of nine pitch shape categories are set. Shall be. For example, in the pitch template 511, the section 514 belongs to the category N-H. At this time, the pitch templates 512 to 514, which are created by correcting the pitch template 511 to a low level, are similarly subjected to pitch shape category classification. And
In the speech segment database, central speech segments of these categories are secured. As a result, the pitch pattern of any sentence can be generated from a finite pitch template.Speech fragments similar to this pitch pattern are secured in the speech fragment database. It is possible to synthesize voices of any pitch pattern such as sentences.

[0021] (Embodiment 4) Next, the implementation of the fourth invention will be described in detail. FIG. 6 is a configuration diagram of the speech synthesis device according to the fourth embodiment. 601 is a reading input unit, 602 is a phoneme chain determining unit, 603 is a prosody determining unit, 604 is a speech unit selecting unit, 605.
, A speech piece connection means, 607, a synthesized speech output means, 608, a pitch template DB (database), 609, a category selection table, and 610, a speech piece database.

First, an intermediate language, which is character information of synthesized contents, is inputted by the reading input means 601. Then, the intermediate language is analyzed by the phoneme chain determination means 602, and a phoneme chain required as a voice segment is determined. The prosody determination means 603 extracts information such as accent and intonation from the intermediate language, and selects a pitch template of a target pitch pattern from the pitch template DB 608. Category selection table 609
Are configured as shown in FIGS. 1 and 3 for each pitch template. The speech segment database 610 is constructed so that speech segments of the category as shown in FIG. 2 are all present for each phoneme chain. As a result, the speech piece selecting means 604 can always select a speech piece having a pitch shape similar to the target pitch pattern. The speech segment selected in this way is transformed by the speech segment transformation unit 605, connected by the speech segment connection unit 606, and output from the synthesized speech output unit 607, thereby constructing a speech synthesis device that suppresses sound quality deterioration due to pitch change. Can be.

Next (Embodiment 5), the fifth implementation of the present invention will be described. FIG. 7 is an example of a category selection table according to the fifth embodiment. In FIG. 7, the categories of the pitch shape are classified into 13 types according to FIG. Thirteen types of pitch shape categories are provided for every phoneme chain, a central voice segment of each category is secured in a voice segment database, and the ID of the voice segment is recorded in a table. By constructing the category selection table and the speech segment database in this way, a speech segment database necessary and sufficient from the viewpoint of pitch pattern can be constructed, so that the size of the speech segment database can be optimized. .

(Embodiment 6) Next, Embodiment 6 corresponding to claim 2 of the present invention will be specifically described. FIG. 8 is a configuration diagram of a speech synthesis device according to the sixth embodiment. In FIG. 8, reference numeral 801 denotes reading input means, 802 denotes phoneme chain determination means, 803 denotes prosody determination means, 804 denotes speech piece selection means, 805 denotes speech piece deformation means,
806 is a speech piece connection means, 807 is a synthesized speech output means,, 80
8 is a pitch template DB with N-mora M-type accent
(Database), 809 is a category selection table for each pitch template of N-mora M-type accent, 810 is a speech piece database, and 811 is a pitch template generating means for sentences.

In the pitch template DB 808, only data whose height has not been corrected, such as the templates 511, 521 and 531 shown in FIG. 5, are recorded. Then, in the category selection table 809, as shown in FIG. 5, the pitch shape category in the case where the pitch template is modified to 1.0 times, 0.9 times, 0.8 times,... Is also recorded. When synthesizing sentences, the sentence pitch template generating means 81 is used.
In step 1, a pitch template DB 808 for a sentence is generated by combining pitch templates with N-mora M-type accents whose pitch pattern has been lowered toward the end of the sentence. Then, a speech piece is selected from the category selection table 809 according to the amount by which the pitch template is corrected low. Thus pitch template D
By configuring the B808 and the category selection table 809, even if the sizes of the pitch template DB 808 and the category selection table 809 are limited, any long text can be synthesized. Therefore, pitch template DB
The size of 808 and the category selection table 809 can be reduced.

[0026]

As described above, according to the present invention, it is possible to guarantee that voice segments resembling the pitch pattern to be controlled exist in the voice segment database, so that it is possible to suppress sound quality deterioration due to pitch change. This has an advantageous effect that the quality of synthesized speech can be improved.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram of a pitch template according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram of category classification of pitch patterns according to the first embodiment of the present invention.

FIG. 3 is a conceptual diagram of a category selection table according to the first embodiment of the present invention.

FIG. 4 is a conceptual diagram of a category classification of a pitch pattern according to the second embodiment of the present invention.

FIG. 5 is a conceptual diagram of a pitch template and a category selection table according to the third embodiment of the present invention.

FIG. 6 is a block diagram of a speech synthesizer according to a fourth embodiment of the present invention.

FIG. 7 is a conceptual diagram of a category selection table according to the fifth embodiment of the present invention.

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

[Explanation of symbols]

201 to 106 Data for each mora position in the pitch template 201 to 205 Pitch shape classification 211 to 214 Pitch pattern category classification 401 to 413 Pitch pattern category classification 511 to 514 Pitch template 521 to 524 Pitch template 531 to 534 Pitch template 601 , 801 reading and input means 602, 802 phoneme chain determination means 603, 803 prosody determination means 604, 804 speech piece selection means 605, 805 speech piece transformation means 606, 806 speech piece connection means 607, 807 synthesized speech output means 608, 808 pitch Template DB 609, 809 Category selection table 610, 810 Voice segment database 811 Text pitch template generation means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-10-116089 (JP, A) JP-A-8-123469 (JP, A) JP-A-11-344997 (JP, A) JP-A 9- 198073 (JP, A) JP-A-10-97291 (JP, A) JP-A-2000-66695 (JP, A) Ryo Mochizuki, Hirofumi Nishimura, Toshimitsu Minowa, Yasuhiko Arai, VCV multi-segment database used for waveform connection synthesis Of the Acoustic Society of Japan, Proceedings of the Acoustical Society of Japan, The Acoustical Society of Japan, September 1998, Autumn 1998, 287-288 (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 13/08

Claims (2)

(57) [Claims] 1. CV, CV / VC, VCV, CV / VC
A waveform superimposition method in which a speech unit of a synthesis unit such as V is selected from a database, the pitch of the selected speech unit is transformed according to a template that is a pattern of the pitch, and these transformed speech units are connected to synthesize speech. In the speech synthesis method, the template, which is the approximate pattern of the pitch, is divided into sections of the synthesis unit, and each section is classified into categories according to the pitch height and shape. When composing a sentence , the speech segment database is configured to
Modify template to lower pitch as
And then combine these modified templates
Use as a whole template
Classify the level to be corrected into several
Which category does the height and shape of the pitch belong to?
A speech synthesis method characterized in that speech pieces are selected in advance according to the category classification . 2. Inputting a voice reading to be synthesized
Input means and CV, CV / VC, V
Phoneme sequence of speech unit of synthesis unit such as CV, CV / VCV
A phonological chain determining means for determining a chain;
Voice segment data that categorizes the pitch height and shape of the
Database and a pitch pattern that is a rough pattern of the pitch of words and sentences.
Pitch template data recording pitch templates
Database and pitch template for pitch control
Means for determining the pitch pattern according to
Switch the template into sections of synthesis units, and
Which category does the pitch height and shape belong to in the section
Category selection table that records information on whether
A speech piece that selects a speech piece according to the category selection table
Selection means and the pitch of the speech piece according to the pitch template
Connects voice-segment deforming means that deforms
Speech unit connection means for synthesizing, and synthesized speech for outputting synthesized speech
Voice output means, and a rough pattern of pitch of words and sentences
Is the length of a word or compound word
Set to, and when synthesizing sentences,
Modify the plate so that it gets lower at the end of the sentence,
A sentence combining these modified pitch templates
Pitch template for writing pitch templates
Providing rate generating means, each pitch template
Category for each level classification that corrects the pitch low in
Voice synthesis characterized by preparing a plurality of reselection tables
Equipment.