JP2001265374A - Voice synthesizing device and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice synthesizing device which can generate a voice message with natural voice quality and rhythm. SOLUTION: As for selected databases 2A, 2B... for voice synthesis, a database feature quantity calculation part 4 calculates the feature quantities of their voice waveforms and then a FUJISAKI model parameter calculation part 5 calculates parameters of a FUJISAKI model. Further, a target rhythm generation part 8 generates target rhythm by using the parameters of the FUJISAKI model. An inputted text is converted by a phoneme connection part 8 into synthesized voice by connecting phonemes selected from the databases 2. Then a rhythm correction part 11 compares the rhythm of the synthesized voice with that of the target voice and corrects the synthesized voice so that they match.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a voice synthesizing apparatus and a recording medium for generating a voice message from an input text or the like, editing the voice message and causing the voice message to be spoken.

[0002]

2. Description of the Related Art A conventional voice message creating / editing device is:
Rule synthesis and waveform superposition are used for the speech synthesis method. The synthesized speech created by these methods has a problem that the prosody is natural, but the voice quality (the degree of likeness of a human voice) is electronic and unnatural. Further, in these systems, it is difficult to change the voice quality of the synthesized speech, and it is not possible to perform differentiation or individualization by voice.

In order to solve the above problem, a speech synthesis method of a speech unit connection type (Japanese Patent Laid-Open No. Hei 10-49193) has been proposed. In this method, a speaker (model) reads aloud a standard text of a language to be spoken and records it.
A phoneme component database is created in which the phonemes are divided into units and the phonemes in the language are covered. When outputting synthesized speech, the phoneme components are recombined according to the text to be read. By using this phoneme parts database, it is possible to obtain a synthesized voice that is close to the real voice of a person who has read out.

A voice synthesized by phoneme connection has a natural voice quality close to that of a real voice, and a plurality of speakers can create a synthesized voice by changing a phoneme component database used at the time of synthesis. . However, this phoneme connection type speech synthesis method has a problem that the prosody of the synthesized speech (the length of the speech and the position of the accent) becomes unnatural.

[0005] As means for changing the prosody of speech,
There are a voice pitch conversion method (Japanese Patent No. 261868) and a voice utterance speed conversion method (Japanese Patent No. 261868). According to these methods, the desired waveform can be changed to an arbitrary prosody by adjusting the waveform to be corrected to the waveform to be a teacher or by manually operating the waveform to be corrected.

However, the former method compares a student's recorded voice with a teacher's recorded voice, and converts the waveform of the student's recorded voice according to the waveform of the teacher's recorded voice. In applications such as synthesizing an arbitrary text and reading it out electronically, it is impossible to prepare a recorded voice or waveform as a teacher for all cases. Also, in combination with voice recognition, etc.,
When a synthesized speech is created in real time, it is impossible to manually adjust the waveform to correct the prosody.

[0007] The current state of speech synthesis is as described above, and it is extremely difficult to achieve both natural voice quality and natural prosody in voice synthesis used in voice message creation / editing devices and text-to-speech devices. Met.

In the case of using a phoneme parts database of a plurality of speakers, when a speaker is switched, the synthesized voice becomes unsmooth or unnatural due to the individuality or characteristics of the speaker. There was something.

[0009]

DISCLOSURE OF THE INVENTION The present invention has been made in view of the above-mentioned prior art, and an object of the present invention is to produce a voice message with natural voice quality and smoothness, particularly with natural voice quality and prosody. A speech synthesizer is provided.

A speech synthesizing apparatus according to the present invention includes a speech synthesis database storing phoneme waveform data relating to a plurality of speakers, and means for extracting features from phoneme waveform data of each speaker stored in the speech synthesis database. And a phoneme combining means for selecting and connecting a phoneme waveform corresponding to the target character from the speech synthesis database, and using the feature according to the selected speaker, the synthesized speech generated by the phoneme combining means. Correction means.

The synthesized voice generated by connecting the phonemes has a characteristic that the voice quality is natural. further,
According to this speech synthesizer, a feature corresponding to a speaker is extracted, and the synthesized speech is corrected using the feature. Therefore, when a plurality of speakers are switched to create a synthesized speech, a different speaker is used. Even if the database is switched to the phoneme speech synthesis database, the synthesized speech is corrected in accordance with the characteristics of the phoneme waveform of the selected speaker, so that the synthesized speech can be generated with natural voice quality and smoothness.

The correcting means in the embodiment of the present invention corrects the prosody of the synthesized speech using the characteristics according to the selected speaker. By correcting the prosody of the synthesized speech by the correction means, it becomes possible to synthesize speech with natural voice quality and prosody according to the selected speaker.

In particular, as an embodiment of the present invention, a speech synthesis database in which phoneme waveform data relating to a plurality of speakers is stored, and a parameter determination for determining parameters of an utterance model from characteristics of each speaker in the speech synthesis database. Means, and means for generating a target prosody at the time of speech synthesis using the parameters determined by the parameter determining means,
Based on the target prosody generated by the target prosody generation means,
Desirably, a phoneme combining means for selecting and connecting a phoneme waveform corresponding to the target character from the speech synthesis database, and comparing the prosody of the synthesized speech with the target prosody by comparing the prosody of the synthesized speech with the target prosody. .

Further, the recording medium according to the present invention includes means for extracting features from the phoneme waveform data of the speaker stored in the speech synthesis database, and selecting a phoneme waveform corresponding to the target character from the speech synthesis database. Using the phoneme coupling means to connect with the above, the feature according to the selected speaker,
Means for correcting the synthesized speech generated by the phoneme combining means. As such a recording medium,
Usually, it is realized in the form of a CD, DVD, MO or the like.

By executing a program stored in such a recording medium, a feature corresponding to a speaker is extracted, and a synthesized speech by phoneme connection is corrected using the feature. Even if the database is switched to a phoneme speech synthesis database, synthesized speech can be generated with natural voice quality and smoothness.

Also, in this recording medium, if the correction means corrects the prosody of the synthesized speech using the feature corresponding to the selected speaker, the correction means naturally adjusts according to the selected speaker. It is possible to synthesize speech with a proper voice quality and prosody.

Particularly, as an embodiment of this recording medium,
Parameter determining means for determining parameters of the utterance model from the characteristics of the speaker in the speech synthesis database; means for generating a target prosody at the time of speech synthesis using the parameters determined by the parameter determining means; Means for selecting and connecting a phoneme waveform corresponding to the target character from the speech synthesis database based on the target prosody generated by the means, and comparing the prosody of the synthesized speech with the target prosody to determine the prosody of the synthesized speech. It is desirable to match the target prosody.

[0018]

FIG. 1 is a diagram showing a configuration of a speech synthesizer 1 according to an embodiment of the present invention. This speech synthesis device 1 includes speech synthesis databases 2A, 2B,..., A speaker database selection unit 3, a database feature amount calculation unit 4,
It is composed of a Fujisaki model parameter calculation unit 5, a text reading unit 6, a phoneme sequence conversion unit 7, a target prosody generation unit 8, a phoneme connection unit 9, a prosody display unit 10, a prosody correction unit 11, a synthesized speech output unit 12, and a storage unit 13. Have been. In addition, a prosody operation unit 14 is provided as needed. The speech synthesizer 1 is configured as a program operating on a computer stored in a recording medium such as a CD, a computer system storing the program in a recording medium such as a CD or a hard disk, or as a dedicated device. . FIG. 2A shows a screen display (user interface) on a computer. The display screen (parent window) 21 includes a text display window 22 for displaying input text, a phoneme notation display window 23 for displaying phoneme notation converted from text, a prosody display window 24, and an operation panel 25. A provided combo box 26 for speaker selection, an utterance button 27, a waveform storage button 28, and the like are provided. FIG. 2B shows a state where the combo box 26 for speaker selection is opened. Hereinafter, the speech synthesizer 1 will be described with reference to FIGS. 1 and 2A and 2B.

The speech synthesis databases 2A, 2B,...
Is a database that holds a speech waveform from which speech is synthesized, character data of phonemes corresponding to the speech waveform and its contents, segments for each phoneme, and prosodic feature parameters of each phoneme. Is stored in Speech synthesis databases 2A, 2B, ...
Is prepared for each different speaker.

The speaker database selection unit 3 selects a speaker to be used for speech synthesis from a plurality of registered speech synthesis databases 2A, 2B,. That is, FIG.
A variety of types of speakers are registered in the combo box 26 for speaker selection as shown in FIG. 3. When the combo box for speaker selection on the screen is opened and a speaker is selected, the voice synthesis for the speaker is performed. The databases 2A, 2B,... Are selected by the speaker database selection unit 3.

The database feature calculation unit 4 selects or designates one of the speech synthesis databases 2A, 2B,... (Hereinafter, the selected speech synthesis database may be referred to as the selection database 2). When done
The feature amount is calculated for the audio waveform in the selection database 2. That is, for the voiced section in the voice waveform extracted from the selection database 2, the basic frequency of each window is calculated by taking an autocorrelation with a fixed window width, and the average value (fo) is calculated as a feature amount for the basic frequency of each window. Calculate mean and standard deviation (fo) std.

The Fujisaki model parameter calculation unit 5 calculates the parameters of the Fujisaki model corresponding to each speaker by using the speaker's speech frequency feature stored in the selection database 2. The Fujisaki model is a model of the fundamental frequency pattern of speech when reading a Japanese declarative sentence.
It is expressed by an equation. Here, Api and Aaj are coefficients, and p
Represents a phrase component and a represents an accent component.

[0023]

(Equation 1)

The Fujisaki model parameter calculation unit 5 uses the features of the speech synthesis database (selection database 2) of each speaker obtained by the database feature calculation unit 4 for parameters dependent on the speaker of this model. decide. In the above Fujisaki model equation (1), Api, Aaj, and Fmin are parameters depending on the speaker. In order to realize natural synthesized speech, it is necessary to change the values of these parameters for each of the speech synthesis databases 2A, 2B,... In accordance with the features of the speech in the selection database 2. Therefore, the Fujisaki model parameter calculation unit 5 calculates the three parameters Api, A
With respect to aj and Fmin, values suitable for the feature amount of the speaker are determined.

That is, the Fujisaki model parameter calculator 5
Uses the fundamental frequency average value (fo) mean and the standard deviation (fo) std of the speech synthesis database 2 selected at the time of speaker selection, and converts the parameters of the Fujisaki model in the database 2 into the following (4) Calculated from Equation (6). Fmin = ln {(fo) mean- (fo) std} ... (4) Api=0.3@-0.42ln [(fo) mean] + 0.42ln [(fo) std] @ + 1.1 ... (5) Aai = 0.7 ｛−0.42ln [(fo) mean] + 0.42ln [(fo) std)｝ + 1.1 (6) As a result, among the speech synthesis databases 2A, 2B,. Features are modeled in the form of Fujisaki models.

The text reading section 6 reads an original sentence (text) to be synthesized and uttered into a memory of a computer. The text is input from a keyboard of a personal computer or sent through a line such as the Internet, and these texts are read into the memory of the computer. The read text is displayed in the text display window 22 as kana-kanji mixed text.

The kana-kanji mixed text input from the text reading section 6 is converted into a phoneme string for synthesis by the phoneme string conversion section 7 and displayed in the phoneme notation display window 23 as a phoneme notation with accent. Then, when the utterance button 27 of the operation panel 25 is pressed, a speech synthesis process is executed.

The target prosody generation unit 8 uses the Fujisaki model parameters Api, Aaj, and Fmin determined by the Fujisaki model parameter calculation unit 5 to generate a target prosody at the time of speech synthesis by applying the Fujisaki model. . The target prosody includes a basic voice frequency pattern and a phoneme duration, and the target prosody generation unit 8 includes a process of generating a basic voice frequency pattern and a process of calculating the duration of each phoneme. In the process of generating the voice fundamental frequency pattern,
Api, A determined by the Fujisaki model parameter calculation unit 5
The values of the three parameters aj and Fmin are actually applied to the expression (1) of the Fujisaki model to generate a speech fundamental frequency pattern to be used as a target during speech synthesis. In the process of calculating the phoneme duration, the duration of the phoneme to be synthesized is calculated using the average value of the phoneme durations in the speech database of each speaker.

The phoneme connection unit 9 selects a database based on the target prosody (fundamental frequency pattern, phoneme duration) generated by the target prosody generation unit 8 for the text to be synthesized input from a keyboard or the like. 2, a phoneme (phoneme waveform) is selected, and an image thereof is formed to create a synthesized speech.
The prosody of the synthesized speech and the prosody of the target speech created in this way are graphically displayed in the prosody display window 24 by the prosody display unit 10.

The prosody correction unit 11 compares the prosody (sound fundamental frequency, phoneme duration) of the synthesized speech generated by the phoneme connection unit 9 with the target prosody calculated based on the Fujisaki model. Is corrected by extending or shortening each phoneme of the synthesized speech waveform so that the synthesized speech waveform matches the target value.

The synthesized speech output unit 12 outputs the synthesized speech created in this way. For example, the output synthesized voice is amplified and converted into voice by a speaker or the like, or the storage unit 1
3 and appropriately stored in a recording medium.

When the waveform save button 28 on the operation panel 25 is pressed, or automatically, the synthesized voice created by the phoneme combining unit 9 and the synthesized voice corrected by the prosody correction unit 11 are stored in the storage unit 13 as a voice waveform. Is done. When the prosody operation unit 14 is provided, the desired prosody can be changed to an arbitrary prosody by operating the target prosody displayed in the prosody display window 24 or the prosody of the actually selected phoneme.

Therefore, according to the speech synthesizer 1, when a synthesized speech is created by a plurality of speakers, the feature amount of the selected speaker can be extracted and modeled by the Fujisaki model. Then, after combining the phonemes extracted from the speech synthesis database to create a synthesized speech, the prosody of the synthesized speech can be corrected by the prosody of this model.
Therefore, according to the voice synthesizing device 1, a voice message with a natural prosody can be created while maintaining the voice quality of each of the voice databases.

[0034]

According to the speech synthesizing apparatus and the recording medium of the present invention, a smooth synthesized speech with natural voice quality can be obtained. In particular, it is possible to utter with natural voice quality and prosody.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a speech synthesis device according to an embodiment of the present invention.

FIG. 2A is a diagram showing a display screen of a speech synthesizer,
(B) is a diagram showing a state in which the speaker selection combo box is opened.

[Explanation of symbols]

 2A, 2B, ... Speech synthesis database 3 Speaker database selection unit 4 Database feature calculation unit 5 Fujisaki model parameter calculation unit 8 Target prosody generation unit 9 Phoneme coupling unit 11 Prosody correction unit 22 Text display window 23 Phoneme display window 24 Prosody display window

Continued on the front page (72) Inventor Toshiyuki Sano 10-Family Todocho Hanazono, Ukyo-ku, Kyoto-shi F-term in Omron Corporation (reference) 5D045 AA01 AA09 9A001 DD15 HH18 KK54

Claims (3)

[Claims] 1. A speech synthesis database storing phoneme waveform data for a plurality of speakers, a means for extracting features from phoneme waveform data of each speaker stored in the speech synthesis database, and a speech synthesis database Phoneme combining means for selecting and connecting a phoneme waveform corresponding to the target character,
Means for correcting the synthesized speech generated by the phoneme combining means using the feature according to the selected speaker. 2. The speech synthesizer according to claim 1, wherein the correction means corrects the prosody of the synthesized speech using the feature according to the selected speaker. 3. A means for extracting features from speaker's phoneme waveform data stored in a speech synthesis database, a phoneme coupling means for selecting and connecting a phoneme waveform corresponding to a target character from the speech synthesis database, Means for correcting the synthesized speech generated by the phoneme combining means using the feature according to the selected speaker.