JP3423276B2 - Voice synthesis method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing method capable of reading arbitrary text information by synthetic voice.

[0002]

2. Description of the Related Art FIG. 1 shows a schematic structure of a speech synthesizer.

The input text containing Japanese kana and kanji characters is subjected to morphological analysis and dependency analysis in the language processing unit 1 and converted into phoneme symbols, accent symbols and the like.

The prosody pattern generation unit 2 uses the phoneme symbol, the accent symbol string, and the part-of-speech information of the input text obtained from the result of morphological analysis, and uses the phoneme duration (voice length DUR ^T ) and the fundamental frequency (voice Pitch (FO ^T ), power of vowel center (loudness POW ^T ) and so on are estimated.

[0005] In the phoneme unit selector 3, estimated phoneme duration DUR ^T, fundamental frequency FO ^T and closest to the power POW ^T vowel center, and phonemes stored in the waveform dictionary 5 (phoneme) The combination of phonemes that produces the smallest distortion when the are connected is selected using DP (dynamic programming).

The speech waveform generator 4 generates speech by connecting the phonemes while converting the pitch according to the selected combination of the phonemes.

FIG. 2 shows the contents of the waveform dictionary 5.
The waveform dictionary 5 includes a phoneme piece storage unit 51 that stores a plurality of phoneme pieces, and an auxiliary information storage unit 52 that stores auxiliary information regarding each phoneme piece in the phoneme piece storage unit 51. The auxiliary information includes the power (POW ^Dic ) of the phoneme piece and the fundamental frequency (FO
^Dic ), duration (DUR ^Dic ), etc.

By the way, the phoneme unit selecting section 3 selects a combination with less distortion among the combinations of phoneme pieces stored in the waveform dictionary 5. The distortion is as follows. is there.

That is, as shown in FIG. 3, u _i-1 ,
u _i and u _{i + 1} are phoneme pieces extracted from the waveform dictionary 5,
Assuming that t _i−1 , t _i , and t _{i + 1} are actually used environments (targets), the distortion with respect to u _i is C _i ^t and C _i.
There is ^c .

Here, C _i ^t is a phoneme piece (u _i ) extracted from the dictionary for the i-th phoneme and the environment actually used.
(Target t _i ) and the distortion. Also, C _i ^c
Is the i-th phoneme (u _i ) and the i−1-th phoneme (u _i ).
_i-1 ) is the distortion that occurs when and are connected. The phoneme unit selection unit 3 connects the phoneme pieces using the dynamic programming (DP method), and outputs C _i ^t and C for all input phonemes.
A combination of the pieces that minimizes the sum C ^all with _i ^c is selected.

C _i ^t is expressed by the following equation 1.

[0012]

[Equation 1]

In equation 1, each variable is defined as follows.

D _POW ^t (t _i , u _i ) is the power (POW ^Dic (i)) of the phoneme piece (u _i ) extracted from the dictionary for the i-th phoneme and the environment (target t) actually used. It is the square of the distance between the power of ( _i ) and the power (POW ^T (i)) and is {(POW ^Dic (i)) − (POW ^T (i))} ² .

W _POW ^t is a weighting factor for D _POW ^t (t _i , u _i ).

D _F0 ^t (t _i , u _i ) is the fundamental frequency (FO ^Dic (i)) of the phoneme piece (u _i ) extracted from the dictionary for the i-th phoneme and the environment (target) to be actually used. It is the square of the distance between the fundamental frequency (FO ^T (i)) of t _i ) and is {(FO ^Dic (i)) − (FO ^T (i))} ² .

W _F0 ^t is a weighting coefficient for D _F0 ^t (t _i , u _i ).

D_DUR ^t(T_i, U_i) Is the i-th phoneme
For phonemes (u_i) Duration
Length (DUR^Dic(i)) and the actual usage environment (target
To t _i) Duration (DUR^T(i)) the distance between
Squared, {(DUR^Dic(i)) − (DUR^T(i))}²When
Become.

W _DUR ^t is a weighting coefficient for D _DUR ^t (t _i , u _i ).

C _i ^c is expressed by the following equation 2.

[0021]

[Equation 2]

In equation 2, each variable is defined as follows.

D_POW ^c(U_i, U_i-1) Is the i-th sound
Element (u_i) Starting power (POW ^DicS(i)) and i-
First phoneme (u_i-1) End power (POW^DicE(i
-1)) is the square of the distance to {(POW^DicS(i))
− (POW^DicE(i-1))}²Becomes

W _POW ^c is a weighting coefficient for D _POW ^c (u _i , u _i-1 ).

D _F0 ^c (u _i , u _i-1 ) is the fundamental frequency (FO ^DicS (i)) of the starting end of the i-th phoneme piece (u _i ) and i
The fundamental frequency (FO at the end of the -1st phoneme unit (u _i-1 )
^DicE (i-1)) is the square of the distance to, {(FO ^DicS
(i))-(FO ^DicE (i-1))} ² .

W _F0 ^c is a weighting coefficient for D _F0 ^c (u _i , u _i-1 ).

D_SPC ^c(U_i, U_i-1) Is the i-th sound
Element (u_i) Beginning spectrum (SPC^DicS(i, j), j = 1
 ~ 16) and the i-1th phoneme piece (u_i-1) End of
Spectrum (SPC^DicEbetween (i-1, j), j = 1 to 16)
It is the square of the distance, and {(SPC ^DicS(i, j))-(SPC^DicE
(i-1, j))}²Becomes

W _SPC ^c is a weighting coefficient for D _SPC ^c (u _i , u _i-1 ).

C _i ^t and C for all input phonemes
The total sum C ^all with _i ^c is represented by the following Expression 3.

[0030]

[Equation 3]

[0031]

By the way, as described above, according to the voice synthesizing method, it is possible to obtain a high quality synthetic voice, that is, a synthetic voice close to natural speech. However, there is a high possibility that a phoneme piece created from a natural utterance may have a phoneme piece such as “smoothness” or “good stagnation” that leads to deterioration in sound quality when actually selected. Although it is preferable to create the waveform dictionary 5 so as not to include such phonemes, it is actually difficult to create the waveform dictionary 5 by removing all phonemes that lead to sound quality deterioration.

It is also conceivable to delete the phonemes that lead to the deterioration of the sound quality after creating the waveform dictionary 5, but in that case, the waveform dictionary 5 needs to be largely modified.

It is an object of the present invention to provide a speech synthesizing method capable of making it difficult to select a phoneme piece of poor quality which leads to deterioration of sound quality as an optimum phoneme piece without making a large modification to the waveform dictionary. To do.

[0034]

In a first speech synthesis method according to the present invention, a plurality of speech units and auxiliary information used for calculating distortion with a target for each phoneme unit are stored in a waveform dictionary. In the phoneme unit selection type speech synthesis method, which selects the combination with the least distortion from the target among the combinations of phoneme units stored in the waveform dictionary, penalty information is added to the auxiliary information of each phoneme unit. The step of adding, the step in which the user listens to the speech synthesis result, and if the quality is poor, the step of causing the user to input the poor quality synthesized speech portion, and the poor quality synthesized speech portion input by the user are input. If this is the case, the phoneme piece is selected as a candidate for the penalty information in the auxiliary information of the phoneme piece corresponding to the poor-quality synthesized speech part. Characterized in that it comprises a step of setting a value that forcibly increases the distortion calculation value of the target when.

In the second speech synthesis method according to the present invention, a plurality of speech units and auxiliary information used for calculating the fitness for the target for each phoneme unit are stored in the waveform dictionary, and the waveform dictionary stores the auxiliary information. A step of adding priority information to auxiliary information of each phoneme unit in a phoneme unit selection-type speech synthesis method that selects a combination having the largest fitness for a target among combinations of stored phoneme units. , If the user hears the speech synthesis result and the quality is poor, the step of causing the user to input the poor quality synthesized speech location, and if the poor quality synthesized speech location input by the user is input, , The priority information in the auxiliary information of the phoneme piece corresponding to the synthesized speech part of the poor quality is selected when the phoneme piece is selected as a candidate. Characterized in that it comprises a step of setting a value that is forcibly reduced calculation value of fitness for Tsu bets.

[0036]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below.

[1] Description of First Embodiment The overall configuration of the speech synthesizer is the same as in FIG.

In the first embodiment, the following points (1),
(2) and (3) are different from conventional ones.

(1) As shown in FIG. 4, penalty information D ^t _pri (u _i ) is added to the auxiliary information of each phoneme piece. The initial value of the penalty information D ^t _pri (u _i ) is 0.
Is.

(2) The penalty information D is added to C _i ^t for calculating the distortion C ^all in the phoneme unit selection unit 3.
Add ^t _pri (u _i ) as a parameter.

That is, C _i ^t is expressed by the following equation 4.

[0042]

[Equation 4]

(3) The user hears the voice synthesis result,
If the quality is poor, the poor-quality synthesized speech portion is input to the speech synthesizer. When a poor-quality synthesized speech portion input by the user is input, the speech synthesis device receives penalty information D ^t _pri (u _i ) in the auxiliary information of the phoneme unit corresponding to the poor-quality synthesized speech location.
Is set to a predetermined value α.

The predetermined value α is, for example, Equation 1
A value of about 100 times the expected maximum value of C _i ^t of is used. Specifically, the maximum value of Equation 1 when an arbitrary number of sentences is input is obtained by an experiment, and the maximum value of 1 is calculated.
A value multiplied by 00 is set as the predetermined value α.

By making the above changes (1), (2) and (3), the penalty information D
^When a poor-quality phoneme piece (u _i ) for which α is set as the value of ^t _pri (u _i ) is selected as a candidate, the distortion C _i ^t between the phoneme piece and the target is determined by the conventional method. Compared with this, it becomes larger by α, and it becomes difficult to select the phoneme piece (u _i ) as the optimum phoneme piece.

According to the above embodiment, when there is a poor quality phoneme piece in the waveform dictionary, the auxiliary information of the phoneme piece is deleted without making a large correction of the dictionary such as deleting the phoneme piece. By making a small modification such as adding penalty information D ^t _pri (u _i ) to
It becomes possible to make it difficult to select a phoneme piece with poor quality.

In the case of a high quality speech synthesizer, since the phoneme piece storage unit in the waveform dictionary stores about 60,000 phoneme pieces, the capacity of the phoneme piece storage unit reaches several tens of MB. , The capacity of the auxiliary information storage in the waveform dictionary is several MB,
It is 1 or less, which is a sufficient capacity of the phoneme piece storage unit. Therefore, it is easier to modify only the auxiliary information storage unit as in the above embodiment. Further, in the conventional method of improving the quality by deleting the phoneme pieces, it is necessary to replace the entire waveform dictionary, but in the method of the above embodiment, the penalty information D ^t _pri (u _i ) is added to the auxiliary information. Therefore, the correction can be made by only changing a part of the waveform dictionary.

It is also possible for the user to improve the quality of synthesized speech by freely deleting low quality phoneme pieces from the waveform dictionary. However, depending on the type of phoneme, all phoneme pieces corresponding to the phoneme piece may be improved. Might be deleted. Then, there is a possibility that a synthesized voice cannot be generated for a sentence including the phoneme.

On the other hand, in the method according to the above embodiment, even if the value of the penalty information D ^t _pri (u _i ) for all the phoneme pieces corresponding to a certain phoneme is set to the predetermined value α, When synthesizing the phoneme,
Since an optimum phoneme piece is selected from the phoneme pieces corresponding to the phoneme, there is an advantage that a synthetic speech can be generated for the phoneme.

[2] Description of Second Embodiment In the first embodiment, in the phoneme unit selection section 3,
Among the combinations of phoneme pieces accumulated in the waveform dictionary,
Although a combination with less distortion is selected, as a phoneme unit selection unit, there is known a phoneme unit selection unit that selects a combination with a large fitness from among combinations of phoneme pieces accumulated in the waveform dictionary.

The fitness S ^all is generally expressed by the following equation 5.

[0052]

[Equation 5]

In equation 5, S_i ^tIs the i-th phoneme
About phoneme pieces (u _i) And actually use
Environment (target t_i) And the similarity between
Is expressed by the following equation 6. Each variable in Equation 6 is an equation
It is the same as the variable in 1.

[0054]

[Equation 6]

In equation 5, S_i ^cIs the i-th
Phonemes (u) selected from the dictionary_i) Starting point
And the phoneme piece selected from the dictionary for the i-1th phoneme
(U _i-1) Indicates the similarity to the end of
It is expressed by Equation 7. Each variable in Equation 7 is a variable in Equation 2
Is the same as.

[0056]

[Equation 7]

In the second embodiment, the following points (1),
(2) and (3) are different from the conventional example in which the phoneme unit is selected using the fitness.

(1) Priority information E ^t _pri (u _i ) is added to the auxiliary information of each phoneme piece. Priority information E
The initial value of ^t _pri (u _i ) is a predetermined value.

(2) The fitness S ^all in the phoneme unit selection unit 3
The S _i ^t for calculating the priority information E
Add ^t _pri (u _i ) as a parameter.

That is, S _i ^t is expressed by the following equation 8.

[0061]

[Equation 8]

(3) The user hears the voice synthesis result,
If the quality is poor, the poor-quality synthesized speech portion is input to the speech synthesizer. When a poor-quality synthesized speech portion input by the user is input, the speech synthesizer inputs priority information E ^t _pri (u _i ) in the auxiliary information of the phoneme unit corresponding to the poor-quality synthesized speech portion. Set the value of to a value smaller than the initial value.

[0063]

According to the present invention, it is possible to make it difficult to select a phoneme piece of poor quality that leads to sound quality deterioration as the optimum phoneme piece without making a large modification to the waveform dictionary.

[Brief description of drawings]

FIG. 1 is a block diagram showing an overall configuration of a speech synthesizer.

FIG. 2 is a schematic diagram showing the contents of a waveform dictionary 5.

FIG. 3 shows two types of distortions C _i ^t and C _i used for selecting a combination of phoneme pieces in a phoneme unit selection unit 3.
It is a schematic diagram for ^{demonstrating c} .

[Fig. 4] Supplement of phoneme pieces corresponding to a synthesized speech portion of poor quality
Penalty information D for auxiliary information ^t _priAdded
It is a schematic diagram which shows.

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 13/06

Claims (2)

(57) [Claims] 1. A plurality of voice units and auxiliary information used for calculating distortion with a target for each phoneme unit are stored in a waveform dictionary, and combinations of phoneme units stored in the waveform dictionary are stored. Among them, in the phoneme unit selection type speech synthesis method that selects the combination with the least distortion with the target, the step of adding penalty information to the auxiliary information of each phoneme unit, the user hears the speech synthesis result, If the quality is poor, the step of prompting the user to input a poor-quality synthesized speech location, and if the poor-quality synthesized speech location input by the user is input, correspond to the poor-quality synthesized speech location. In the penalty information in the auxiliary information of the phoneme piece to be set, the distortion calculation value with the target is forcibly increased when the phoneme piece is selected as a candidate. Speech synthesis method characterized in that it comprises a step of setting a so that value. 2. A plurality of voice units and auxiliary information used to calculate the fitness for a target for each phoneme unit are stored in a waveform dictionary, and combinations of phoneme units stored in the waveform dictionary are stored. Among them, in the phoneme unit selection type speech synthesis method that selects the combination with the highest fitness for the target, the step of adding priority information to the auxiliary information of each phoneme unit, the user hears the speech synthesis result. , If the quality is poor, the step of prompting the user to input a poor quality synthesized speech location, and if the poor quality synthesized speech location input by the user is entered, The priority information in the auxiliary information of the corresponding phoneme piece is forced to the calculated value of the fitness for the target when the phoneme piece is selected as a candidate. Speech synthesis method characterized by comprising the step, a setting a value that is small.