JP3133347B2 - Prosody control device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a prosody control device for a rule speech synthesizer.

[0002]

2. Description of the Related Art As a prosody control system of a ruled speech synthesizer, there is, for example, one disclosed and filed by the present applicant as a "prosody control system" in Japanese Patent Application Laid-Open No. 63-46498. The prosody is controlled using a multivariate analysis technique such as quantification class I or multiple regression analysis. As another example of the prosody control method, as having been conceived also by the applicant, by selectively used depending on the situation at that time the regression coefficient sets and speakers, different speaker characteristics Ya In some cases, the tone can be easily expressed.

[0003]

Conventionally, a synthesized speech by a regular speech synthesizer has a serious disadvantage that it lacks naturalness and makes the listener feel uncomfortable. Such problems are mainly caused by the control of prosody. Although recent research has gradually reduced such discomfort, it has become possible to synthesize natural synthesized speech. But,
Humans do not always speak in a fixed prosody pattern, and the prosody pattern naturally differs from speaker to speaker.
As described above, the prosody of the conventional synthesized voice has little change, and it is a reality that listening for a long time is still painful.

When a prosody pattern is generated by the above-described conventional example, variable selection for analyzing and predicting the prosody pattern becomes an important problem. In particular, unnecessarily increasing the number of variables in the regression analysis may lead to a decrease in prediction accuracy for unknown inputs. The conventional method uses a single set of variables for any speaker or tone, so it is necessary to use variables of the greatest common divisor, which makes detailed analysis and prediction difficult. .

[0005]

According to the present invention, a speech unit parameter sequence prepared in advance is read out according to an input symbol string representing phonemes and prosody, and the speech unit parameter sequences are connected according to a speech parameter combination rule. In a rule speech synthesizer that adds a prosody by calculating a phoneme duration, a pitch pattern, an amplitude pattern, etc. according to an input symbol string expressing a phoneme / prosody by a control rule, a plurality of variables are extracted from the input symbol string by a variable extraction rule. the variable extraction rule with extracts of variables to select an optimum set of variables for different speakers and utterance form from these variables
And the optimized variable set and the optimal variable
Optimum obtained from the set and the input symbol string and voice data
Category mapping means for storing various coefficient sets , and the variable sets are stored in accordance with different speakers or utterance forms.
It has a variable storage selection means you choose from stored by categories mapping means, selected in this variable storage selection means
The variable set and the corresponding coefficient set
From the memory stored by the
Synthesized speech by number set, coefficient set and input symbol string
To generate a prosody pattern .

[0006]

By using the category mapping means and the variable memory selection means, it is possible to finely control the prosody pattern of the synthesized speech, which was conventionally uniform, including the variable selection.

[0007]

An embodiment of the present invention will be described with reference to the drawings. 1 and 2 show a configuration example of a prosody control device in a rule speech synthesizer. First, as shown in FIG. 1, a labeling 2 is connected to the template audio unit 1, and a category mapping unit 3 as a category mapping unit is connected to the labeling 2. A regression analysis unit 4 is connected to the category mapping unit 3. Further, as shown in FIG. 2, a regression prediction unit 5 is connected to the category mapping unit 3.

In this case, the category mapping unit 3
Has a function of extracting a plurality of variables from an input symbol string 6 according to a category determination rule 7 as a variable extraction rule, and selecting an optimum variable set 8 for each different speaker or utterance form from these variables. Further, here, the category determination rule 7, the optimized variable set 8, and the regression coefficient set 9 are stored, and these are selectively used according to the purpose of use of the synthesized speech and the speaker characteristics. , A variable storage selecting means (not shown) is provided.

In such a configuration, in this embodiment,
Take phoneme duration as an example of a prosody pattern,
Consider the case where quantification class I is used for prosody pattern analysis and prediction. Hereinafter, it is assumed that a voice data set uttered by a human is given in advance, and this is called a template voice here. The configuration and operation of the present embodiment will be described separately for an analysis for analyzing and learning the characteristics of the prosody pattern included in the template voice and a synthesis for generating a prosody pattern based on the result.

First, the operation at the time of analysis will be described. FIG. 1 shows the configuration at the time of analysis. A phoneme boundary position is determined by a labeling unit 2 from a template voice uttered by a human obtained from the template voice unit 1,
The duration of each phoneme is measured and used as the target variable. On the other hand, in the category mapping unit 3, an explanatory variable is obtained from the label string given by the labeling unit 2 by the category determination rule 7. The explanatory variable in the quantification class I is generally called a factor item and is a qualitative one to which category the sample falls. What is important here is which factor item is to be adopted, and further, what kind of category is to be used. Therefore, in the present embodiment, the variable selection optimized for the template speech to be analyzed is performed without taking the greatest common denominator variable set. As a variable selection method, a known method such as a forward selection method, a backward elimination method, or a sequential method using an variance ratio, a degree of freedom-adjusted contribution ratio, or the like as an evaluation function can be used. Regression analysis is performed by the regression analysis unit 4 using the optimal variable set selected in this way, and a regression coefficient set 9 is obtained.

Next, the operation at the time of synthesis will be described. FIG. 2 shows a configuration at the time of synthesis. It is assumed that an input symbol string 6 corresponding to a label string input at the time of analysis is obtained from a text or a concept to be uttered. From the input symbol string 6, the category mapping unit 3 performs mapping to an appropriate category by using the category determination rule 7 used at the time of analysis and the optimized variable set 8. Then, using the regression coefficient set 9 obtained at the time of the analysis, a phoneme duration pattern corresponding to the input symbol string 6 is calculated. By synthesizing the speech in this way, a synthesized speech that reproduces the characteristics of the given template speech can be obtained. In addition, various template voices such as a speaker and a utterance form are separately analyzed, and the category determination rule 7, the optimized variable set 8, and the regression coefficient set 9 are stored by the variable storage selection means and appropriately used, so that various types are obtained. By appropriately setting the prosody pattern 10, it is possible to reproduce a variety of sounds.

[0012]

According to the present invention, a speech unit parameter sequence prepared in advance is read in accordance with an input symbol string representing phoneme and prosody, the speech unit parameter sequence is connected by a speech parameter combination rule, and a phoneme is controlled by a prosody control rule.・ In a rule speech synthesizer that adds prosody by calculating phoneme duration, pitch pattern, amplitude pattern, etc. according to an input symbol string expressing prosody, multiple variables are extracted from the input symbol string by a variable extraction rule. From among these variables, an optimal variable set is selected for each of different speakers and utterance forms, and the variable extraction rule and the optimized
Variable set and this optimal variable set and the input
Describe the optimal coefficient set obtained from the symbol string and audio data.
Provide category mapping means to remember different speakers and
The variable set is assigned to the category mapping method according to the voice form.
Have a variable storage selection means you choose from stored by the step
And the variable set selected by the variable storage selecting means and
The corresponding coefficient set is recorded by the category mapping means.
Take these variables sets and coefficient sets out of
Prosody pattern of synthesized speech by
Since it is generated, it is possible to finely control the prosody pattern of the synthesized speech, which was conventionally uniform, including variable selection, and thereby synthesize a natural speech. .

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration at the time of voice analysis according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration at the time of speech synthesis according to an embodiment of the present invention. 3 Category mapping means 6 Input symbol string 7 Variable extraction rule 8 Variable set

Continuation of the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 13/00-13/08 G10L 21/00-21/06

Claims (1)

(57) [Claims] 1. A speech unit parameter sequence prepared in advance is read in accordance with an input symbol string representing phoneme / prosodic, said speech unit parameter sequence is connected by a speech parameter combination rule, and a phoneme / prosody is represented by a prosody control rule. In a rule speech synthesizer that adds a prosody by calculating a phoneme duration, a pitch pattern, an amplitude pattern, and the like according to an input symbol string, a plurality of variables are extracted from the input symbol string according to a variable extraction rule. And selecting the optimal variable set for each different speaker or utterance form, and the variable extraction rule and the optimized variable
Set and the optimal variable set and the input symbol string and
Mosquitoes <br/> categories mapping means for storing an optimum coefficient set obtained from a fine voice data provided, the different speakers and utterance form
The variable set is stored by the category mapping means according to
Is has a variable storage selection means you choose from the, this variant
The variable set selected by the number storage selection means and the corresponding
Several sets are stored in the category mapping means.
From these variables and coefficient sets and inputs
Generates a prosody pattern of a synthesized speech using a symbol string
Prosody control device, characterized in that had Unishi.