JPH06332490A - Generating method of accent component basic table for voice synthesizer - Google Patents

Abstract

PURPOSE:To faithfully express accent information of actual human voice and to realize smoothness in an accent control during a synthesis. CONSTITUTION:For every accent control point (seven points for each data), an averaged value of all data belong to a group is obtained in a first step S1. Employing the averaged value, a standard deviation is obtained in a second step S2. Only using the data which exist within a + or - range of the averaged value and the standard deviation, an averaged value is again obtained in a third step S3 and the results are registered in a table for every group in a fourth step S4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention realizes a pitch pattern closer to human occurrence by providing a plurality of pitch target values per phoneme in controlling the pitch of synthesized speech and controlling the accent width in multiple stages. The present invention relates to a method of creating an accent component basic table of a voice synthesizing device which is referred to when a basic pattern of an accent component is composed in a regular speech synthesizing device.

[0002]

2. Description of the Related Art As shown in FIG. 3, a speech synthesizing system using a rule synthesizing method uses a morphological analysis by a Japanese processing unit 2 for a text input from a text input unit 1 to find a position of a pose and a phrase of a word or phrase. The phonetic conversion and accent addition are performed by cutting and referring to a dictionary. For example, when the input text is "Today's weather is nice", the Japanese processing result is shown in the table below.

[0003]

[Table 1]

For the text resulting from this processing, in the intonation control unit 3, the phrase pattern calculation unit 3A obtains a phrase component (in pauses) obtained from the number of mora (consonant + minimum unit of sound represented by vowels) included in the text. The pitch of a sound when a person is caught in one breath is calculated), the accent component (the pitch of the sound of each word) is calculated by the accent pattern calculation unit 3B, and the respective components are superposed as shown in FIG. The intonation control pattern is calculated. The phrase component gradually decreases from a high pitch at the beginning when a person speaks to a low pitch due to a decrease in subglottic pressure. Also,
As described above, the accent component may be given one pitch target value for each analysis unit and linearly interpolated between them, or may be given three pitch target values for each analysis unit and linearly interpolated between them.

In the latter accent component calculation, target values P _{1 to} P ₃ are given to the categories shown in Tables 2 and 3 below for each CV unit (consonant + vowel) and V unit (vowel), and the category indicated by * is added. The pitch is linearly interpolated from the preceding and following data.

[0006]

[Table 2]

[0007]

[Table 3]

The speech synthesizer 4 adjusts the pitch of each syllable according to the intonation control pattern provided by the intonation processor 3, and outputs the response of the articulatory filter from the source wave pattern and the articulatory filter parameter associated with each syllable. As a synthetic voice.

As described above, the conventional intonation control system imparts intonation to the synthesized voice by synthesizing the phrase component according to the number of moras contained in the input text and the accent component for each motor. Of these, the accent component has a plurality of target pitches per mora, and the pitch change during that time is determined by interpolation processing.

[0010]

Since the pitch of the accent component is simply determined from the height of the accent,
It was easy to get a uniform pitch change and mechanical synthetic voice. In addition, since the connection between syllables and syllables and syllables and syllables is not considered at all, the change in accents before and after the mora may become unsmooth, resulting in unnatural synthetic speech as syllables or sentences as a whole. There was a problem.

The present invention has been made in view of the above circumstances, and performs averaging processing of accent patterns of a plurality of data for each of various grouped environments, creates a table, and averages from a large number of occurrences. Provides a method for creating a basic table of accent components for a speech synthesizer that extracts a valued accent pattern to more accurately represent the accent information of a human-generated voice and realizes smoothness in accent control during synthesis. The purpose is to do.

[0012]

In order to achieve the above object, the present invention relates to an input control pattern for a Japanese language input text, which is based on the number of mora in the phrase of the input text. The pitch target value determined for each mora in the phrase of the input text is corrected and interpolated according to the accent environment to superimpose accent components to obtain an intonation control pattern of the input text. A basic accent pattern table that stores the patterns in a table is obtained, a basic accent pattern corresponding to the accent environment of the input mora is obtained from the table, and a basic accent pattern generation processing unit that corrects the pitch of the mora is obtained. The processing result of this processing unit is used for the Japanese phrase processing In the speech synthesis method of the rule synthesizing method, which comprises a correction processing unit that obtains the accent pattern of the accent component by correcting according to the number of mora and the mora position, the basic accent pattern table groups the environments in which various mora are uttered, The averaging process is performed on each accent control point for each group, and then the processing result is registered in the table for each group.

In the second aspect of the invention, the averaging process is performed by averaging Xa for all data belonging to the group for each accent control point.
After obtaining ve, the standard deviation SD is obtained, and thereafter, the average value is obtained again using only the data existing within the range of the average value Xave ± 3SD.

[0014]

When the accent basic table is created, the accent data registered for each mora environment is calculated.
At this time, the patterns of the data groups belonging to the environment group are averaged, and an appropriate average accent pattern is registered in the table. This improves continuity between data and improves the naturalness of the entire synthetic speech.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. In FIG. 1, S ₁ is an accent control point (for example, 7 points for each data: j = 1 to 1) for each group when the environment in which various mora are uttered is grouped.
In the first step when performing the averaging process for 7),
This first step S ₁ is, for each accent control point (7 points for each data), average value Xave ₁ for all data [X ₁ (j) to X _N (j)] belonging to the group.
Is obtained by the equation (1).

[0016]

[Equation 1]

[0017] The average value Xave ₁ obtained in the first step S ₁
Based on the above, the standard deviation SD is obtained in the second step S ₂ .
This standard deviation SD is obtained using the equation (2).

[0018]

[Equation 2]

After the standard deviation SD is obtained, the process proceeds to the third step S ₃ and the average value Xave is obtained again using only the data (N ′ pieces) existing within the range of Xave1 ± 3SD. Expression (3) is an expression for obtaining the average value.

[0020]

[Equation 3]

The results obtained as described above are registered in the table for each group. This step is the fourth.
In this way, when creating the accent basic table, the accent data registered for each mora environment is first calculated. After the calculation, the patterns of the data groups belonging to the environment group are averaged, and a more appropriate average accent pattern is registered in the table. If registered in the table, the continuity between data can be maintained and the naturalness of the synthetic speech as a whole can be improved by using the following configuration diagram.

FIG. 2 is a block diagram when the accent pattern obtained in FIG. 1 is registered in the table 7. The same parts as those in FIG. 3 are designated by the same reference numerals. In FIG. 2, the accent pattern generation processing unit 5 is provided with the number of moras in a section and the accent pattern data (see Table 1) of the moras as the Japanese processing result of the input text. The accent pattern generation processing unit 5 processes this data by the basic accent pattern generation processing unit 6 and the basic accent pattern table 7 and the correction processing unit 8 which will be described in detail later.

The basic accent pattern generation processing unit 6 is
Processing for reading the basic accent pattern corresponding to the accent environment of the input data from the table 7 is performed. This basic accent pattern table 7 is determined by taking into consideration the types of sounds before and after the mora, accent patterns, etc. when determining the value of the accent component of a mora in a sentence when performing rule composition. This is determined by referring to the values of accent components for determining the accent components for various environments in which mora appears, respectively.

[0024]

As described above, according to the present invention,
Based on a speech database generated from a large number of utterance examples, create an accent basic table to refer to when deciding accent basic components, and generate a pitch pattern for speech synthesis based on this to generate a human voice. When the basic pattern is created when the pitch pattern of No. 1 is reflected in the synthesized sound, the table is created by averaging the accent patterns of a plurality of data for each of various grouped environments. By extracting a more average accent pattern from the utterance, it is possible to more faithfully represent the accent information of the real voice uttered by a human and to realize smoothness in accent control during synthesis.

[Brief description of drawings]

FIG. 1 is a flow chart showing an embodiment of the present invention.

FIG. 2 is a configuration diagram to which FIG. 1 is applied.

FIG. 3 is a block diagram of a conventional speech synthesis apparatus using a rule synthesis method.

FIG. 4 is a diagram showing an intonation processing mode.

[Explanation of symbols]

S ₁ ... 1st step S ₂ ... 2nd step S ₃ ... 3rd step S ₄ ... 4th step

Claims (2)

[Claims] 1. For input text processed in Japanese,
An intonation control pattern of a phrase component is obtained from the number of mora in the phrase of the input text, and a pitch target value defined for each mora in the phrase of the input text is corrected and interpolated in accordance with the accent environment to superimpose the accent component on the input text. To obtain a basic accent pattern table in which the basic accent patterns for each mora accent environment are stored and stored, and a basic accent pattern corresponding to the input mora accent environment is obtained from the table. A basic accent pattern generation processing unit for correcting the pitch of the mora is obtained, and a correction processing unit for correcting the processing result of the processing unit by the number of mora in the phrase cut and the mora position which are processed in Japanese to obtain an intonation pattern of the accent component. Speech Synthesis Method with Rule Synthesis Method In the basic accent pattern table, the environments in which various mora are uttered are grouped, the averaging process is performed for each accent control point for each group, and the processing result is registered in the table for each group. A method of creating an accent component basic table for a speech synthesizer characterized by the above. 2. The averaging process obtains an average value Xave for all data belonging to a group for each accent control point, then obtains a standard deviation SD, and thereafter, the average value Xave.
2. A method of creating an accent component basic table of a speech synthesizer according to claim 1, wherein the average value is obtained again only by using the data existing within the range of ave ± 3SD.