JPH04191898A - Text speech synthesizer - Google Patents

Abstract

PURPOSE:To determine the optimum characteristic parameter according to vocal sound environment and utterance speeds without forming complicated rules by using a fuzzy computing means which makes computation to the fuzzy set calculated by 1st, 2nd fuzzy set calculating means. CONSTITUTION:The 1st fuzzy set calculating means 17 calculates the 1st fuzzy set indicating the characteristic parameter influenced by the preceding sound element and the 2nd fuzzy set indicating the characteristic parameter influenced by the succeeding sound element. The 2nd fuzzy set calculating means 18 calculates the fuzzy set indicating the change rate from the target value of the characteristic parameter influenced by the duration time of the sound element. The fuzzy computing means 19 makes fuzzy computation to the fuzzy sets calculated by the 1st, 2nd fuzzy set calculating means 17, 18 and determines the characteristic parameter of the vocal sound. The optimum characteristic parameters are obtd. according to the vocal sound environment and utterance speeds without forming the complicated rules.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a text-to-speech synthesizer that synthesizes speech from arbitrary character strings.

[Prior Art] Conventional text-to-speech synthesis devices that use vocal tract parameters such as formant frequency and bandwidth as phoneme feature parameters use formant parameters extracted from natural speech in advance as they are, or use the influence of preceding and preceding phonemes. Speech was synthesized by applying predetermined rules and modifying parameters depending on things like speaking speed.

FIG. 9 shows formant extraction by the conventional text-to-speech synthesizer described above. Formant extraction as shown in FIG. 9(B) is performed for each frame from the natural speech shown in FIG. 9(A).

Next, for each phoneme, a frame that most prominently represents the characteristics of that phoneme is selected, and the formant of the selected frame is used as a representative formant for synthesizing that phoneme.

The representative formant is a very important value, especially in the case of vowels, and a speech synthesizer synthesizes speech by repeating the same value (of the representative formant) within the duration of the vowel.

Even in the case of consonants, such as nasal consonants, the voice may be synthesized by repeating the representative formant for the duration.

If synthesis cannot be performed using the representative formant, the formant value extracted for each frame is used as is.

Furthermore, at the boundary between phonemes, the formant value from the formant of the preceding phoneme to the formant of the subsequent phoneme is calculated by using the connection rule.

Connection rules include, for example, performing linear interpolation on the assumption that there is a straight line transition between the representative formant of the preceding phoneme and the representative formant of the succeeding phoneme, or repeating the representative formant within the duration of the preceding phoneme and connecting it to the representative formant of the subsequent phoneme. In the continuation section of a phoneme, linear interpolation is performed by considering it as a straight line transition from the representative formant parameter of the preceding phoneme.

FIG. 10 shows an example of analysis of the synthesized speech synthesized in this manner. Figure 10 (A) shows the waveform of the synthesized speech, Figure 10 (
B) shows the formants of the synthesized speech.

``Problem to be solved by the invention 2 In the conventional speech synthesis device described above, in order to improve the quality of synthesized speech, the influence of the phonological environment such as preceding phonemes and subsequent phonemes, and the speaking rate (duration of phonemes) is minimized. Therefore, it is necessary to select the optimal feature parameters for each phoneme.
Therefore, there is a problem in that a huge number of rules must be created and characteristic parameters selected in advance must be modified.

In view of the above-mentioned problems with conventional speech synthesis devices, an object of the present invention is to provide a speech synthesis device that can obtain optimal feature parameters according to the phonetic environment and speech rate without creating complicated rules. .

[Means for Solving the Problems] The present invention provides a method for calculating a first fuzzy set representing a feature parameter influenced by a preceding phoneme and a second fuzzy set representing a feature parameter influenced by a subsequent phoneme. a second fuzzy set calculation means for calculating a fuzzy set representing a rate of change from a target value of a feature parameter affected by the duration of a phoneme;
and fuzzy calculation means for calculating phoneme feature parameters, the fuzzy calculation means performing fuzzy calculations on the fuzzy sets calculated by the first fuzzy set calculation means and the second fuzzy set calculation means. This is achieved by a text-to-speech synthesizer configured as follows.

[Operation] The first fuzzy set calculation means calculates a first fuzzy set representing feature parameters influenced by the preceding phoneme and a second fuzzy set representing feature parameters influenced by the subsequent phoneme. , the second fuzzy set calculation means calculates a fuzzy set representing the rate of change from the target value of the feature parameter affected by the duration of the phoneme, and the fuzzy calculation means calculates the first and second fuzzy set calculation means. A fuzzy operation is performed on the fuzzy set calculated by the means to obtain phoneme feature parameters.

[Example] Hereinafter, a text-to-speech synthesis device of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the text-to-speech synthesis apparatus of the present invention.

The text-to-speech synthesizer shown in FIG. 1 includes a character string analysis section 11,
It is composed of a word dictionary 12, a rule control section 13, a rule file I4, a speech synthesizer 15, and a feature parameter file 16.

Next, the operation of the text-to-speech synthesizer shown in FIG. 1 will be explained focusing on each of the above-mentioned components.

When an arbitrary character string is input, the character string analysis unit 11 analyzes the syntax of the input character string and determines the intonation pattern of the entire character string.

The character string analysis unit l) further refers to the word dictionary 1.2,
The words included in the input character string are searched, and the accent and phoneme sequence of each word in the character string are determined to determine the phoneme sequence and accent pattern of the character string.

The character string analysis unit 11 outputs the intonation of the entire character string, the phonetic sequence of the character string, and the accent pattern thus determined to the rule control unit 13.

The feature parameter file 16 outputs parameters representing features of phonemes to the rule control unit 13.

The rule file 14 outputs to the rule control unit 13 phoneme control rules for connecting the feature parameters output from the feature parameter file 16 and prosody control rules for controlling each prosody. This prosodic control rule includes the duration of phonemes, which will be described later.

The rule control unit 13 refers to the feature parameters from the feature parameter file 16 and the phonological control rules and prosody control rules from the rule file 14 to determine the intonation pattern of the entire character string input from the character string analysis unit 11,
Based on the phonological sequence and accent pattern of the character string, and the feature parameters modified by the procedure described below,
Speech synthesizer 15 generates parameters necessary for speech synthesis.
Output to.

The speech synthesizer 15 generates synthesized speech based on the parameters inputted from the rule control section 13, and outputs the rule synthesized speech corresponding to the human-generated character string. In order to obtain the parameters, components as shown in FIG. 2 are provided.

The configuration of the rule control section 13 will be explained below with reference to FIG.

The rule control unit 13 includes a first fuzzy set calculation unit 17 that calculates a first fuzzy set representing feature parameters influenced by the preceding phoneme and a second fuzzy set representing feature parameters influenced by the subsequent phoneme. , a second fuzzy set calculation unit 18 that calculates a fuzzy set representing the rate of change from the target value of the feature parameter affected by the duration of the phoneme, a first fuzzy set calculation unit 17 and a second fuzzy set calculation unit 17
The fuzzy calculation unit 19 performs fuzzy calculations on the fuzzy set calculated by the fuzzy set calculation unit 18 to obtain feature parameters.

The fuzzy calculation unit 17 of plan 1 calculates the membership function representing the first fuzzy set based on the appearance frequency of the feature parameter for each preceding phoneme, and calculates the membership function representing the first fuzzy set based on the appearance frequency of the feature parameter for each subsequent phoneme. Membership function calculation unit 2 that calculates a membership function representing the fuzzy set of 2
It is equipped with G.

The second fuzzy set calculation unit 18 expresses the duration of the phoneme as a fuzzy set, and further expresses the rate of change from the target value of the feature parameter for each fuzzy variable of the duration as a fuzzy set, and calculates the given duration. The apparatus includes a fuzzy inference section 21 that performs fuzzy inference to calculate a fuzzy set of change rates.

The fuzzy inference 1[2+ includes a target value calculation unit 22 that calculates a target value of a feature parameter based on the probability of occurrence of the feature parameter in a fuzzy set of feature parameters obtained from natural speech.

Next, a procedure for obtaining the modified feature parameters by the rule control unit 13 described above will be explained.

Here, the representative formant will be specifically explained as a feature parameter. The representative formant refers to the formant frequency and bandwidth that best represent the characteristics of phoneme influenced by the preceding phoneme, the following phoneme, and the duration of the phoneme.

The membership function of the representative formant is determined by the following procedure. However, although one of the formant parameters will be explained below, all parameters can be determined by the same procedure.

The target value of each phoneme is determined by the step target value calculation unit 22.

Vocalization data is statistically processed for each phoneme under various conditions, and the fuzzy value with the highest probability of occurrence is targeted using fuzzy probability.

The target calculation procedure will be explained below.

In the frequency domain, the entire region is assumed to be Ω, and frequencies such as the formant frequency and bandwidth of each data are assumed to be variables ω.

Next, for the phoneme Xi, the probability p (F) that the parameter occurs in the fuzzy variable F representing the frequency domain is determined. As shown in Figure 3, if the membership function of the fuzzy variable F is μF(ω), then P (F) = f, μF(ω)・dP = f, μF(ω)・p(ω)・dω p (ω)=n (ω)/N FCΩ, ω(Ω, μF (ω) Ω→[0,1) From this, the probability of occurrence P(F) of the fuzzy variable F is determined.

Here, N represents the total number of data, and n(ω) represents the frequency of data at frequency ω.

As a result, the target value Ti of the phoneme Xi is calculated using the following formula: P (Ti)=max (P (F));T
+=FFCΩ The membership function of the target value Ti is μTi(ω).

Step 2 The second fuzzy set calculation unit 18 calculates a fuzzy set representing the displacement of the representative formant from the target according to the duration of the phoneme.

The displacement of the representative formant obtained in step 1 from the target T is statistically determined according to the duration.

Since the target T is represented by a fuzzy variable, the displacement to be determined is a fuzzy number.

Displacement Di (1) from the target Ti when the duration of the phoneme Xi is short (the membership function is μDi (1)
) is μDi(1)=μL (1) ・μT, (1, dω)
Calculated by

Here, μF(1) is the grade whose duration l belongs to the fuzzy variable group, μT, , (1°dω) is μ
The frequency when the duration of Ti(ω) is l is 1Ti−f
,,l/1Ti-Tj The membership function normalized by l, Tj is the target of the preceding (following) phoneme, f
ll is the observed value of the representative formant when the preceding (following) phoneme is Xj.

Furthermore, the displacement Di(L) of the representative formant of the phoneme Xi belonging to the duration d from the target is calculated from the following equation.

Di(L)=\7Di(:l) FIGS. 4A to 4C show the membership function and calculation process of the calculated displacement D1 of the representative formant from the target.

Step 3: The membership function calculation unit 20 calculates representative formants classified based on the preceding phoneme as a fuzzy set.

Based on the probability of occurrence of the representative formant influenced by the preceding phoneme Xj of the phoneme Xi, a membership function μfi (o<i<p; p is the formant degree) of the representative formant is determined.

Step 4: The membership function calculation unit 2Q calculates representative formants classified based on the subsequent phonemes as a fuzzy set.

Based on the probability of occurrence of a representative formant influenced by the phoneme Xj following phoneme X1, the membership function μbi of the representative formant (0<i<p; p is the formant order)
seek.

Membership function μf1 obtained in steps 3 and 4 above
, μb1 is shown in FIG.

Using the fuzzy set of representative formants of phonemes obtained in the above procedure (steps 1 to 4), representative formants of phonemes are determined and speech is synthesized according to the procedure shown below.

Step 5: The first fuzzy set calculating unit 17 extracts fuzzy sets Ff and Fb of representative formant parameters using the preceding phoneme and the following phoneme as keys.

Step 6 The fuzzy inference unit 21 performs fuzzy inference based on the duration to find a displacement function from the target of the formant parameter affected by the duration.

In step 2 above, the membership function μDi representing the displacement from the target is calculated for each fuzzy variable of duration.

Here, we calculate the grade for the duration of the actually given duration expressed by a fuzzy variable.
member subfunction μDO corresponding to the given duration
Find i as a displacement function from the target.

The calculation method will be explained with reference to FIG.

The duration is expressed by three fuzzy variables: Long, Middle, and 5hort. Given a given duration, apply the following three fuzzy rules:

(A) If Long, then DO is DL.

(B) If it is Middle, then DO is DM.

(C) If Short, then DO is DS.

However, DL, DM, and DS each have a long duration.
, Middle, and 5hort are membership functions representing the displacement from the target.

Figures 6 (A), (B), and (C) represent the above three fuzzy rules using membership functions.If the duration is given, the magnitude for the fuzzy variable can be found.

When the duration is given as a fuzzy value, α is the grade obtained as a result of the max-min operation between the above fuzzy variables, and when the duration is given as a crisp value, the grade obtained as a result of the max-min operation with the fuzzy variable is The grade is α,
Apply the rules above. FIG. 6 shows a case where the duration is given fuzzy.

The elements of the displacement functions DL, DM, and DS are 0 for the target of the target phoneme and 1 for the target of the preceding or subsequent phoneme.
It represents the distance on the parameter between the two corresponding to.

When considering the formant frequency, it becomes a continuous value in which each target frequency corresponds to 0 and 1.

Fuzzy inference calculates the grade α using the given duration and the antecedent part of the rule, calculates the displacement function obtained from each rule by truncating the displacement function by α, and calculates the displacement function obtained from each rule. The target displacement function DO is obtained by performing OR combination. Figure 6 CD)
shows the displacement function DO obtained as above.

In the fuzzy inference process described above, the great α is calculated using the given duration and the antecedent of the rule, the displacement function obtained from each rule is calculated by truncating the displacement function at α, and the displacement function obtained from each rule is calculated. Although the target displacement function DO was obtained by ORing the displacement functions obtained, the method for obtaining the displacement function DO is not limited to this, and for example, a method as shown in FIG. 7 may be used. can.

That is, as shown in FIG.
), (B), and (C), the grade α1. determined by the antecedent part of each rule. α5. The displacement function obtained from each rule is calculated by multiplying the displacement function by α0.

Then, to synthesize the displacement functions obtained from each rule, marginal sum is used instead of the above-mentioned OR synthesis.

Here, the marginal sum is an operation expressed by the following equation.

μc=(4μ8 to μ)Δ1 In addition to the above-mentioned OR combination and marginal sum, various methods can be used to synthesize the membership functions. FIG. 7(D) shows the displacement function D obtained as above.
Indicates O.

Step 7 The fuzzy calculation unit 19 synthesizes the membership functions obtained in steps 5 and 6 above.

The representative formant fuzzy sets Ff, Fb influenced by the preceding and subsequent phonemes and the fuzzy set DO representing the displacement from the target due to the influence of duration are AND-synthesized as shown in FIG. 8 to obtain the representative formant. Get Fm.

Here, in order to synthesize DO, Ff, and Fb, the frequency axis must be converted. Therefore, the membership function μDO of DO is linearly expanded or contracted so that 0 corresponds to the target value of the phoneme of interest and 1 corresponds to the target value of the preceding or succeeding phoneme.

Step 8 Defuzzify the fuzzy set of representative formants.

In the example shown in FIG. 8, the representative formant was determined by the centroid method, which is the most widely used method for defuzzification. However, defuzzification is not limited to this method.

Through the above procedure, it is possible to obtain a representative formant that is compatible with the phoneme environment and also takes into account the duration of the phoneme.

Depending on the phoneme, parameters change over time,
In some cases, it is difficult to express phoneme using one representative formant.

In the above case, the present invention can be applied by considering a plurality of representative formants over time.

Synthetic speech can be created by connecting the representative formants obtained in the above procedure. When connecting, smoothing is performed as necessary.

In the above embodiment, the representative formant was used as an example of the phoneme feature parameter, but this is just an example, and the phoneme feature parameter 9 may be a parameter obtained by linear predictive analysis etc. It is also possible to use feature parameters.

[Effects of the Invention] A first fuzzy set calculating means for calculating a first fuzzy set representing a feature parameter influenced by a preceding phoneme and a second fuzzy set representing a feature parameter influenced by a subsequent phoneme; a second fuzzy set calculation means for calculating a fuzzy set representing a rate of change from a target value of a feature parameter affected by the duration of a phoneme; and a fuzzy calculation means for calculating a phoneme feature parameter; Since the fuzzy operation means is configured to perform fuzzy operations on the fuzzy sets calculated by the first fuzzy set calculation means and the second fuzzy set calculation means, the fuzzy operation can be performed without creating complicated rules. Optimal feature parameters can be determined according to the phonological environment and speech rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the text-to-speech synthesizer according to the present invention, FIG. 2 is a block diagram showing details of the rule control section 13 configuring FIG. 1, and FIG. Graphs showing membership functions of fuzzy variables. Figures 4 (A) to (C) are graphs of membership functions representing the displacement of representative formants according to the duration of a phoneme. Figure 5 shows the influence of the environment. Graphs showing membership functions of received feature parameters, Figures 6 (A, ) to (D
) is a graph showing an example of fuzzy inference for duration time, Figure 7 (A) to (D) is a graph showing another example of fuzzy inference for duration time, and Figure 8 (A) to (C) is a graph showing another example of fuzzy inference for duration time. , a graph showing an example of composition of membership functions, FIGS. 9(A) and (B) are graphs showing an example of formant extraction of natural speech, and FIGS. 10(A) and (B) are graphs showing an example of formant extraction of natural speech. It is a graph showing an example of analysis of synthesized speech by a synthesizer. 11...Character string analysis section, 12...Word dictionary, 13.
...Rules side @ Edict, 15...Rules file, I5...
Speech synthesizer! 6...Feature parameter file, 17
... first fuzzy set calculation section, 18 ... second fuzzy set calculation section, 19 ... fuzzy operation section, 2Q.
... Menpan sub function calculation section, 21 ... Fuzzy inference section, 22 ... Terkerat value calculation section. ；Main B's Tardrid μ Figure 8 Fm

Claims (1)

[Claims] a first fuzzy set calculation means for calculating a first fuzzy set representing a feature parameter influenced by the preceding phoneme and a second fuzzy set representing a feature parameter influenced by the subsequent phoneme; It is equipped with a second fuzzy set calculation means for calculating a fuzzy set representing a rate of change from a target value of an affected feature parameter, and a fuzzy calculation means for calculating a phonological feature parameter, the fuzzy calculation means , a text-to-speech synthesis device configured to perform fuzzy operations on fuzzy sets calculated by the first fuzzy set calculation means and the second fuzzy set calculation means.