JPH07319495A - Synthesis unit data generating system and method for voice synthesis device - Google Patents

Abstract

PURPOSE:To obtain a synthesized voice having a desired voice property by using voice data of plural human beings having different voice properties in a voice synthesis device. CONSTITUTION:Synthesis unit data obtained from voices of plural human beings are made to be stored as a data-base 24. When a text and voice wanted to be synthesized into the synthesized voice is inputted, synthesis unit data nearest synthetically in reagard to a phonemic environment, a power, a pitch and a voice property are selected with respect to respective voice synthesis units composing the text. Next, whether the voice property of the speaker of selected synthesis unit data is different from a voice property wanted to be synthesized more than a certain degree or not is checked and in the case the difference is greater than the degree, the voice property of the selected synthesis unit data is converted into the same voice property as the voice property wanted to be synthesized by executing a voice transformation to synthesis unit data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention generally relates to a speech synthesizing device for automatically reading a digitized text, which is utilized for proofreading a prepared sentence, reading by a visually impaired person, and the like. , Improvement of the technique for generating synthesis unit data.

[0002]

2. Description of the Related Art In a conventional speech synthesizer, in order to output a synthesized speech of good quality, it is necessary to provide a sub-category in the speech synthesis unit in consideration of the phonological environment before and after the synthesis unit. For example, "Asahi" / asahi / becomes / a // sa // hi / when written in synthetic units, but the synthetic unit / asah
The phoneme data used is sandwiched between the preceding phoneme / a / and the following phoneme / hi / similar to i / / sa /. For example, "Asahikawa" / asahikawa / / sa / has the same phonological environment as / asahi / / sa / and can be used to synthesize the word / asahi /. However, since a large number of subcategories are provided for all the synthesis units, a large amount of voice data necessary for creating the voice synthesis unit is required. Further, if there is a difference between the parameters such as the pitch and the power of the voice required for the word to be synthesized and the parameters included in the synthesis unit data, the quality of the synthesized voice is significantly deteriorated. Therefore, even if the synthesis units have the same phonological environment, it is necessary to have several types with different powers and pitches, and a large amount of data is needed to create the synthesis units.

On the other hand, a technique for converting the voice quality of voice data has been reported. For example, Shin Abe "Voice Conversion Technology", IEICE Technical Report, Vol.93, No.427, pp69 to pp75, introduces a voice quality conversion method that focuses on the source characteristics and vocal tract characteristics of speech.

[0004]

The speech synthesizer requires a large amount of speech data for creating a synthesis unit because of the reason described in the previous section. Therefore, it is possible to create a synthesis unit by using multiple human voice data recorded in various scenes and join them together to perform voice synthesis, but then the voice quality of each individual will be different. In addition, the quality of synthesized speech is deteriorated. Therefore, it is necessary to record the voice data for creating the synthesis unit from a single person for a long time, which requires a great deal of labor for the recording work.

Therefore, an object of the present invention is to select the voice data to be used as a synthesis unit from a plurality of human voices, connect them, and perform voice synthesis. The goal is to get close synthetic speech.

[0006]

A synthesis unit data generation method of the present invention is a synthesis unit data holding unit for storing synthesis unit data obtained from voices of a plurality of speakers, a synthesis unit to be used for voice synthesis, and voice quality. Input the information that specifies, and select the synthesis unit data closest to the specified synthesis unit from the synthesis unit data holding unit, and the voice quality of the speaker of the selected synthesis unit data is specified. And a voice quality conversion unit for performing voice quality conversion to bring the voice quality of the selected synthesis unit data closer to the specified voice quality.

[0007]

Function: Synthesis unit data obtained from the voices of a plurality of speakers is previously stored in the database. When a synthesis unit to be used for voice synthesis and a voice quality are designated, first, synthesis unit data closest to the designated synthesis unit is selected from the database. Next, it is checked how much the voice quality of the speaker of the selected synthesis unit data differs from the specified voice quality. If the voice quality of the selected unit differs from the specified voice quality by more than a predetermined level, the voice quality of the synthesis unit data is approximated to the specified voice quality. The conversion is done.
In this way, it is possible to obtain the desired unit synthesis data of a single voice quality from the voice data of various speakers having different voice qualities.

[0008]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 shows the overall structure of an embodiment of a sound quality conversion type speech synthesizer according to the present invention.

First, the structure of the apparatus will be described. This sound quality conversion type speech synthesizer inputs a kanji-kana mixed sentence from the input terminal 11 and designates a voice quality to be synthesized (gender, age group, voice pitch, etc., hereinafter referred to as target voice quality). The preprocessing unit 12 linguistically analyzes this input sentence,
A phonological symbol sequence of words forming a sentence, a word accent type, and a sentence pattern of an input sentence are obtained. The selection criterion parameter setting unit 13 determines a synthesis unit (hereinafter, referred to as a target synthesis unit) necessary for speech synthesis and a phonological environment (from the phonological symbol sequence of the obtained word, the accent type of the word, and the sentence pattern of the input sentence). = Preceding phoneme, subsequent phoneme) and predetermined parameters (= power, pitch).

The synthesis unit data selection unit 14 holds a plurality of synthesis unit data obtained from a plurality of human voices in advance for each synthesis unit, and the phonological environment of the target synthesis unit passed from the selection reference parameter setting unit 13. And select synthesis unit data that best matches the parameter and the target voice quality. Furthermore, it is checked how different the voice quality of the selected synthesis unit data is from the target voice quality. If the voice quality is different to some extent or more, the synthesis unit data is converted into a voice quality similar to the target voice quality.

The synthesizing unit data transforming unit 15 transforms the selected synthesizing unit data so as to match the parameters of the target synthesizing unit. The synthesis unit data splicing unit 16 splices the modified synthesis unit data to synthesize the word data. The output terminal 17 converts the synthesized word data into voice and outputs it.

Next, the structure of the synthesis unit data selection unit 14, which is a feature of this speech synthesis apparatus, will be described in detail with reference to FIG.

In FIG. 2, the likelihood calculation unit 21 performs the target synthesis at these points based on the synthesis unit name of the target synthesis unit, the phonological environment, the parameters, and the target voice quality passed from the selection criterion parameter setting unit 13. The identification number of the composite unit data that best matches the unit (hereinafter referred to as the data ID) is
It is determined by referring to the composition unit data information table 22 and the personal codebook storage unit 25.

Here, the composition unit data information table 22
As shown in FIG. 3, for all synthesis unit data stored in the system, the synthesis unit name, phonological environment (preceding phoneme, subsequent phoneme), parameter (power, power,
Pitch), a speaker identification number (hereinafter referred to as a speaker ID), and a data ID are recorded. Further, the personal codebook storage unit 25 stores a speaker codebook in which the speaker IDs and information representing the characteristics of the voice quality of all the speakers of the synthesis unit data stored in the system are recorded. Has been done.

The processing of the likelihood calculation section 21 will be described in detail. First, the phonological environment and parameters of each synthesis unit data are extracted from the synthesis unit data information table 22, and each of them is extracted from the personal codebook storage section 25. Information about the voice quality of the speaker of the synthesis unit data is extracted. Next, the phonological environment, parameters and target voice quality information of the target synthesis unit passed from the selection criterion parameter setting unit 13 and the synthesis unit data extracted from the synthesis unit data information table 22 and the personal codebook storage unit 25 previously. The phonological environment, parameters, and voice quality information are substituted into the evaluation function shown below to obtain the distance D between the target synthesis unit and each synthesis unit data.

D = W1 ・ Pre (s1, t1) + W2 ・ Post (s2, t2) + W3 ・ Po
w (s3, t3) + W4 ・ Pit (s4, t4) + W5 ・ Person (s5, t5) where s1 to s5 are the phonological environment of the target synthesis unit, parameter and target voice quality information, and t1 to t5 are , Phonological environment of synthesis unit data, parameter and voice quality information, Pre () is a distance calculation function for a preceding phoneme, Post () is a distance calculation function for a subsequent phoneme, Pow () is a distance calculation function for power, Pit
() Is a distance calculation function regarding pitch, Person () is a distance calculation function regarding voice quality information, and W1 to W5 are weighting factors. This evaluation function calculates the distance between the target synthesis unit and the synthesis unit data for each of the preceding phoneme, the subsequent phoneme, the power, the pitch, and the voice quality, and considers the influence degree on the quality of the synthesized speech in each distance. Multiplied by the weighting factor
It is the sum of them.

The likelihood calculating section 21 calculates the distance D by
This is performed for all composition unit data stored in the system, and the distance D is the smallest, that is, the degree of coincidence (likelihood).
The synthesis unit data having a high level is determined to be used for speech synthesis, and the data ID of the synthesis unit data, the speaker ID, and the distance (degree of coincidence) regarding the voice quality obtained in the above calculation process are set in the voice quality conversion unit 23. hand over.

Based on the data ID and the speaker ID passed from the likelihood calculation section 21, the voice quality conversion section 23 extracts the actual synthesis unit data specified by this from the individual synthesis unit database 24. Here, in the individual synthesis unit database 24, synthesis unit data obtained from the recorded voice of each speaker is accumulated for each speaker. Subsequently, the voice quality conversion unit 23 checks whether or not the degree of coincidence of the voice qualities passed from the likelihood calculation unit 21 is lower than a predetermined threshold, and if it is low (when the voice qualities are considerably different), the designated voice quality is determined. The voice quality of the extracted synthesis unit data is converted so that the voice quality is almost the same as that of.

As the voice quality conversion method, for example, a method using a speaker codebook in consideration of vocal tract characteristics described in the above "speech conversion technology" is adopted. That is, from the personal codebook storage unit 25, the codebook of the speaker corresponding to the speaker ID (hereinafter referred to as the original codebook).
And the codebook of the speaker having a voice quality that matches the target voice quality information (hereinafter referred to as the target codebook), and the codebook mapping from the original codebook to the target codebook is performed to obtain the voice quality of the synthesis unit data. To the target voice quality.

There are various other known voice quality conversion methods, which may be adopted. For example,
The method described in “Voice conversion considering sound source characteristics” by Kasuya and Yang (Academic Society of Japan, 1-6-18, pp225-226, 1991) can be adopted.

By the above processing, it is possible to synthesize a voice having a desired voice quality by using voice data recorded by a plurality of speakers. Therefore, it is not necessary to record a large amount of voices from one person, and it is possible to collect and reuse various types of voices already recorded.

The present invention can be implemented in various modes other than the above-mentioned embodiments. For example, in the above-described evaluation function, by appropriately increasing the weighting coefficient W5 for the score of the degree of coincidence regarding the voice quality, it is possible to prevent the synthesis unit data of speakers having greatly different voice qualities from being selected. Thereby, the voice quality conversion unit 23 can be omitted.

The synthesizing unit data selecting unit 14 applies this as a synthesizing unit database creating apparatus to create a synthesizing unit database composed of synthesizing unit data having voice qualities of one speaker from voice data of a plurality of speakers. It is also possible to create. When this database is used, in the configuration of the voice synthesizer, the speaker ID can be omitted from the synthesis unit data information table, and the voice quality conversion unit and the personal codebook storage unit are also omitted. be able to. However, when selecting synthesis unit data, considering the degree of matching (distance) of voice qualities (speakers) calculated at the time of database creation, the quality was greatly reduced due to voice quality conversion at database creation. It is better to introduce this distance into the evaluation function in the description in the composition unit data information table because there is no risk of selecting composition unit data.

[0025]

According to the present invention, voice data recorded by a plurality of people are accumulated as synthesis unit data, synthesis unit data close to the synthesis unit desired to be used for synthesis are selected, and this and the voice quality to be synthesized are selected. When the difference is more than a certain amount, the voice quality of the selected synthesis unit data is converted to the target voice quality, so that the synthesized voice of the desired voice quality can be obtained by using the voice data of a plurality of humans. There is no need to record a large amount of voice data from humans, which is effective in reducing the recording labor.

Further, since the voice data of a plurality of persons already recorded can be reused, it is effective in reducing the work of creating the synthesis unit database.

[Brief description of drawings]

FIG. 1 is a block diagram showing the overall configuration of an embodiment of a voice quality conversion type speech synthesizer of the present invention.

FIG. 2 is a block diagram showing a detailed configuration of a synthesis unit data selection unit in FIG.

FIG. 3 is a diagram showing an example of a composition unit data information table in FIG.

[Explanation of symbols]

 11 Input Terminal 12 Pre-Processing Section 13 Selection Criteria Parameter Setting Section 14 Synthesis Unit Data Selection Section 15 Synthesis Unit Deformation Section 16 Synthesis Unit Joining Section 17 Output Terminal 21 Likelihood Calculation Section 22 Synthesis Unit Data Information Table 23 Voice Conversion Section 24 Individual Synthesis unit database 25 Individual codebook storage

Claims (2)

[Claims] 1. A synthesis unit data holding unit for storing synthesis unit data obtained from voices of a plurality of speakers, and inputting information designating a synthesis unit desired to be used for voice synthesis and voice quality, and the designated synthesis unit. And a synthesis unit data selection unit that selects the synthesis unit data closest to the above from the synthesis unit data holding unit, and whether or not the voice quality of the speaker of the selected synthesis unit data differs from the specified voice quality by a predetermined degree or more. However, if different, a voice quality conversion unit that performs a voice quality conversion to bring the voice quality of the selected synthesis unit data closer to the specified voice quality, a synthesis unit data generation method for a voice synthesis apparatus, . 2. A process of designating a synthesis unit to be used for speech synthesis and a voice quality, and a degree of coincidence with the designated synthesis unit from synthesis unit data obtained from voices of a plurality of speakers accumulated in advance. And the process of selecting the highest synthesis unit data, check whether the voice quality of the speaker of the selected synthesis unit data is different from the specified voice quality by a predetermined degree or more. And a step of performing voice quality conversion to bring the voice quality closer to the specified voice quality, the synthesis unit data generation method for a voice synthesis device.