JPH09179577A - Rhythm energy control method for voice synthesis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent voice quality from deteriorating. SOLUTION: A judgement part of a step S4 judges whether or not a rhythm whose fundamental frequency is determined is a consonant and when it is judged that the rhythm is the consonant, it is supplied to a consonant energy control part of a step S5. The consonant energy control part controls the energy of the consonant by making use of data in a table TB for consonant energy. When it is judged that the rhythm is a vowel, it is supplied to a vowel energy control part of a step S7. The vowel energy control part controls the vowel by a quantization class-I system. Then the consonant and vowel are inputted to a rhythm continuance control part of the said step S6, where after rhythm continuance (length of voice) is controlled by using a rhythm continuance quantization class-I coefficient data base DB3, they are passed through a waveform editing part of a step S8; and a voice is synthesized and inputted to a synthesized voice file of a step S9.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a phoneme energy control method in arbitrary Japanese rule speech synthesis.

[0002]

2. Description of the Related Art FIG. 4 is a block diagram showing a schematic configuration of a ruled speech synthesis system. In FIG. 4, a kanji / kana mixed sentence (sentence) input to a text input unit 1 is sent to a Japanese processing unit 2. give. The Japanese processing unit 2 converts the given sentence into a phoneme symbol string with reference to a built-in Japanese dictionary. This phoneme symbol string is input to the phoneme control unit 3 which determines phoneme parameters (fundamental frequency → pitch, pitch → amplitude → phone size, phoneme duration length → phone length) of a phoneme. , Prosodic parameters are generated based on the phoneme symbol string. The prosody parameter is given to the speech synthesizer 4 after the target value is determined for each phoneme based on the data stored in the database of each fundamental frequency, amplitude, and phoneme duration that are not shown. The voice synthesizer 4 refers to a database (not shown) to obtain a voice while realizing a desired prosody parameter.

In the regular speech synthesis system configured as described above, when the phoneme control unit 3 determines the phoneme parameters, factors such as the preceding phoneme, subsequent phoneme, accent environment, phoneme type, etc. of the phoneme are considered. Must-have. Conversely, all these factors can explain the prosodic parameters.

In the phoneme energy control, a method from the input phoneme symbol string to be uttered to the energy determination of a certain phoneme will be described. For this, a method using quantification type I which is one of multivariate analysis is known. This technique is
A target quantitative variable (energy) is calculated based on a qualitative variable (phonological environment) that is a qualitative explanation factor, and is formulated by the following equations (1) and (2). In the following equation, the factor item of the qualitative variable of the i-th data is j,
The category (classification of each item) to which it belongs is k, and the category quantity (quantity “coefficient” given to the category) is a _jk
When

[0005]

[Equation 1]

Here, δ is a variable defined as follows.

[0007]

[Equation 2]

In order to predict the value {y _i } of the quantitative variable by the least squares method, the category quantity {a _jk } is determined so as to satisfy the following expression (2).

[0009]

(Equation 3)

In order to obtain the category quantity using the above method, the qualitative variables (b) to (f) of the energy database described below are input, and the fundamental frequency value of (a) is used as the teacher signal. I want it.

Energy database: One data in the database is composed of energy in a phoneme with a phoneme symbol string and environmental data for explaining it, and is as shown in (a) to (f) below. ing.

(A) Fundamental frequency value (b) In-phrase position (c) In-sentence position (d) Prefix phoneme type (e) Phoneme type (f) Postfix phoneme type Each phonological energy can be estimated by deriving the qualitative variables (b) to (f) in the above database and inputting them into a linear form consisting of category quantities. Since this control method targets the quantitative variables as the objective variables and the qualitative variables as the explanatory variables, the controlled object can be applied not only to the energy but also to the fundamental frequency and the phoneme duration.

[0013]

[Problem to be Solved by the Invention] Human voice utterance, which is one of the measures of the quality of synthesized voice obtained by the above-mentioned voice synthesizing method, is influenced by a prosody parameter. In order to realize it, it is preferable that the prosody parameter of the synthetic speech is closer to the prosody pattern of the real voice uttered by a human. Therefore, as the prosody parameters learned by the quantified type I model, those obtained by analyzing the real voice when a human naturally speaks are used. However, there are many parameters that fail the analysis because the analysis method has not been established yet. In particular, the learning data has a possibility of including a pattern that is not a prosodic pattern of a real voice, because the learning data remarkably appears with respect to the prosody parameter.

By the way, in the quantified type I model, the independence of each category is assumed, and the number of linear coefficients is the sum of the number of categories of each explanatory variable, which is very small. The estimation accuracy is determined by the quality of the learning data. The good or bad here is that, assuming that the partial correlation coefficient of the explanatory variable with respect to the objective variable has a high value to some extent, and the explanatory variable sufficiently explains the objective variable, the data existing in each category is How much
Is it accurately distributed?

Below, example sentences and variables for explaining the energy of the phoneme are shown.

[0016] "Hello, here, interpreting phone, is an international conference secretariat.", "/ MOSHIMOSHI / KOCHIRA / TU-Y A KUD
EXWA / KOKUSAIKAIGIJIMUKYOKUDESU / ”In addition, underline A
Indicates the phoneme.

(A) fundamental frequency → value estimated by the fundamental frequency control unit, (b) position in phrase → located at beginning of phrase, (c) position in sentence → located at beginning of sentence , (D) front phoneme type → semivowel or nasal sound, (e) relevant phoneme type → A, (f) postphoneme type → unvoiced plosive consonant.

Here, the energy estimation of vowels and consonants in the quantified type I method will be compared. As the voice data that is the basis of the vowel energy database and the consonant energy database, “ATR115” sentence utterance data was used. 6 vowels for vowel energy estimation, 1 for consonant energy estimation
Experiments were performed on two consonants. The two experimental conditions are only the difference in the number of categories of the phoneme type, and the difference in the number of categories of the front phoneme type and the postphoneme type due to the phoneme difference. Tables 1 and 2 show the results of the vowel energy estimation experiment and the consonant energy estimation experiment, respectively.

[0019]

[Table 1]

[0020]

[Table 2]

When comparing the above experimental results, the mean squared error is 2.142 dB in vowel energy estimation, which is about a human discrimination threshold, but in consonant energy estimation,
It is extremely large at 4.373 dB. From this experiment, there is a problem that the quantified type I method cannot be used for consonant energy estimation.

The present invention has been made in view of the above circumstances. By adopting a table system in the consonant energy estimation part, an extremely large energy estimation can be prevented.
An object of the present invention is to provide a phoneme energy control method in speech synthesis that prevents deterioration of sound quality.

[0023]

In order to achieve the above object, the first aspect of the present invention is to convert a sentence from a text input unit into a phoneme symbol string in a Japanese processing unit and then convert the phoneme symbol string. After being processed by the phonological symbol sequence processing unit, and input to the fundamental frequency control unit to obtain the pitch of the sound, it is determined whether the phoneme is a consonant or a vowel, and when it is a consonant, the consonant energy control unit It controls the loudness of the consonant using the data from the consonant energy table, and when it is a vowel, it is input to the vowel energy control unit and the fundamental frequency for each phoneme,
After controlling the loudness of the vowels by performing the quantified type I system processing consisting of the position in the phrase, the position in the sentence, the prephonology, the phoneme concerned, and the postpositional phoneme, the consonants and the vowels are phoneme duration control section. It is characterized in that the waveform is edited by inputting into the input to obtain the sound length and then the synthesized voice is obtained.

A second aspect of the present invention is characterized in that the consonant energy table is classified into a front phoneme, a corresponding phoneme, and a back phoneme.

[0025]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a flowchart of a consonant energy control method for energy control showing an embodiment of the present invention. In FIG. 1, step S1 is a phoneme symbol string file, and the phoneme symbol string called from the file of step S1 is After being processed by the phoneme symbol string processing unit in step S2, it is input to the fundamental frequency control unit in step S3 for determining the fundamental frequency of a certain phoneme. The basic frequency control unit determines the basic frequency when the basic frequency FNN is used.
The data of the coefficient database DB1 is used.

Whether or not the phoneme for which the fundamental frequency has been determined is a consonant or not is determined by the determination unit in step S4,
If it is determined that the sound is a consonant, it is supplied to the consonant energy control unit in step S5. The consonant energy control section controls the energy of the consonant (sound volume) using the data in the consonant energy table TB. This table TB is categorized by the front phoneme, the phoneme, and the post phoneme, and the human tunes each phoneme.

If the phoneme is judged to be a vowel in step S4, it is supplied to the vowel energy control section in step S7. In this vowel energy control unit, vowels are controlled by the quantification type I method. Here, the quantification type I method will be described. In this method, a human voice is analyzed off-line in advance, and a fundamental frequency, a position in a phrase, a position in a sentence, a prepositional phoneme, the phoneme, and a postpositional phoneme (an explanatory variable of a quantified type I model) for each phoneme. Further, the energies are obtained and these are used as learning data. This learning data is input to the quantification type I model, the obtained category quantity is stored, and the energy quantification type I coefficient database DB2 is stored.
And

When the speech synthesizer is operating, the value of the estimation explanatory variable is calculated from the phoneme symbol string (step S
2) is input to the vowel energy control unit in step S7 to obtain the desired energy of the vowel.

On the other hand, if the phoneme is a consonant, the quantification I
Since the estimation accuracy is not good in the similar method, it is controlled by the consonant energy control unit in step S5 as described above.

The consonants and vowels obtained at steps S5 and S7 are input to the phoneme duration control unit at step S6, where the phoneme duration is calculated using the quantified class I coefficient database DB3 for phoneme duration. After the time length (sound length) is controlled, it goes through the waveform editing unit in step S8, is then voice-synthesized, and is input to the synthesized voice file in step S9. In the waveform editing unit of step S8, the speech unit data of the speech unit database DB4 and the step S8 are used.
The phoneme symbol string from the second phoneme symbol string processing unit is processed in step S1.
The waveform is edited using the segment data selected by the segment data selection unit of 0.

Next, a case in which the energy of all phonemes is estimated by the quantified type I method and a case in which the energy of consonants is estimated by the table method and the energy of vowels is estimated by the quantified type I method are described as "one week. Only (I SHU KA X BA KA R
2 and 3 show the comparison results using the sentence example "I)". 2 and 3, the horizontal axis represents time, that is, a phoneme symbol string, and the vertical axis represents energy values.

In particular, the energy of the phoneme "B" (encircled in FIG. 2) in the case of the quantification type I method of the energy of all phonemes shown in FIG. 2 is the energy of the "X" of the pre-phoneme. Has taken an extremely large value for. When this is heard with a synthetic voice, since the phonological "B" is a voiced consonant of a burst type, it sounds very strange. On the other hand, when the energy of the consonant shown in FIG. 3 is controlled by the table method and the energy of the vowel is controlled by the quantified type I method, the phoneme “B” (from FIG. Since the energy is relatively smoothly transferred to the energy surrounded by the circle), it sounds as a smooth sound with no discomfort even in the case of synthetic speech. In this way, it is possible to prevent the sound quality from being deteriorated by estimating the energy of the consonant by the table method and the energy of the vowel by the quantified type I method.

[0033]

As described above, according to the present invention,
By adopting the table method for the consonant energy estimation part, it is possible to prevent the estimation of extremely large energy and prevent the sound quality deterioration. Further, since the consonant energy table is classified only by the three parameters of the prefix phoneme, the phoneme, and the suffix phoneme, there is an advantage that it can be easily created. Furthermore, a system can be constructed with a simple system such as a table system for consonants and a quantified type I system for vowels. In addition, since the quantified type I method is used for the vowel part, by having a plurality of energy quantified type I coefficient data (obtained by acquiring learning data of a plurality of humans), There are advantages such as enabling energy control corresponding to the voice.

[Brief description of the drawings]

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

FIG. 2 is a characteristic diagram when the energy of all phonemes is estimated by the quantification type I method.

FIG. 3 is a characteristic diagram when the energy of a consonant is estimated by a table method and the energy of a vowel is estimated by a quantification type I method.

FIG. 4 is a block configuration diagram of a speech synthesizer.

[Description of Codes] S1 ... Phonological symbol sequence file S2 ... Phonological symbol sequence processing unit S3 ... Basic frequency control unit S4 ... Consonant / vowel determination unit S5 ... Consonant energy control unit S6 ... Phonological duration control unit S7 ... Vowel energy control Part S8 ... Waveform editing part S9 ... Synthetic voice file S10 ... Element data selection part TB ... Consonant energy table DB1 ... FNN coefficient database for fundamental frequency DB2 ... Quantification type I coefficient database for energy DB3 ... Quantification for phoneme duration Type I coefficient database DB4 ... Speech element database

Claims (2)

[Claims] 1. A Japanese processing unit converts a sentence from a text input unit into a phoneme symbol string, the phoneme symbol string processing unit processes the phoneme symbol string, and then inputs the phoneme symbol string to a fundamental frequency control unit. After obtaining the pitch of the sound, it is determined whether the phoneme is a consonant or a vowel, and when it is a consonant, it is input to the consonant energy control unit and the volume of the consonant is controlled using the data from the consonant energy table. In the case of a vowel, the vowel energy is input to the vowel energy control unit and, for each phoneme, a quantification I consisting of a fundamental frequency, a position in a phrase, a position in a sentence, a prefix phoneme, the phoneme, and a suffix phoneme I
After controlling the loudness of the vowels by performing the type processing,
A phonological energy control method in speech synthesis, characterized in that a consonant and a vowel are input to a phoneme duration control unit to obtain a sound length and then waveform editing is performed to obtain a synthetic speech. 2. The method of controlling phonological energy in speech synthesis according to claim 1, wherein the consonant energy table is classified into a prephonic phoneme, a corresponding phoneme, and a postural phoneme.