JP3078074B2 - Basic frequency pattern generation 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 method for generating a fundamental frequency pattern used for speech synthesis.

[0002]

2. Description of the Related Art A speech synthesizer for inputting character information, converting it into speech, and outputting the speech is expected to be a technology that can replace recording / playback speech synthesis because there is no restriction on the output vocabulary. I have. In this type of synthesizer, a technique for generating a fundamental frequency (pitch) pattern of vocal cord vibrations representing voice accents and intonation is a very important elemental technique for obtaining a natural synthesized sound.

[0003] Since the pitch pattern of speech is strongly affected not only by the accent type of each word but also by the sentence structure, meaning, etc., various fundamental frequency patterns extracted from actual speech are prepared and combined. It is difficult to achieve this by using the entire sentence pattern, and appropriate modeling is indispensable.

Several methods have been proposed as models for generating a fundamental frequency pattern of speech, but a fundamental frequency pattern on a logarithmic axis is a frame corresponding to a gentle descent (intonation) from the beginning to the end of a sentence. Is expressed as the sum of a noise component and an accent component corresponding to a local undulation (accent), and the phrase component is an approximation that is a response of a critical damping quadratic linear system to an impulse-like phrase command. In general, a model formulated based on an approximation that the accent component is a response of a critical damping quadratic linear system to a step-like accent command is widely used (Kiyoshi Hirose, Hiroya Fujisaki, Tsune Kawai, Mikio Yamaguchi: "Synthesis of Sentence Speech Based on Fundamental Frequency Pattern Generation Process Model", Transactions of the Institute of Electronics, Information and Communication Engineers, '89 / 01 Vol.
-A No. 1).

FIG. 5 is a block diagram showing a conventional fundamental frequency pattern generation model. In this model, the logarithmic fundamental frequency In F0 (t) is given by the following equation as a function of time t.

[0006]

(Equation 1)

Here, F _min is the base frequency, A _pi is the size of the i-th phrase command in the text, A _aj is the size of the j-th accent command in the text, and I is the frame size in the text. The number of accent commands, J is the number of accent commands in one sentence, T _0i is the start time of the i-th _phrase command, and T _1j and T _2j are _j , respectively.
The start and end times of the second accent command.

Gp (t) and Ga (t) are an impulse response function of the phrase control mechanism and a step response function of the accent control mechanism, respectively, and within the range of t ≧ 0, the following equation (2):
(When t <0, Gp (t) = Ga (t) =
0).

Gp (t) = αt · exp (−αt) (2) Ga (t) = Min [1− (1 + βt) · exp (−βt), θ)] (3)

Here, α and β are constants which determine the response speed of the phrase control mechanism and the response speed of the accent control mechanism, respectively, where α = 3.0 and β = 20.0. Use values. Θ is the upper limit value of the accent component, usually θ = 0.
9 mag.

In human utterances, even if uttered flatly, the pitch is high at the beginning of the utterance, and thereafter the pitch naturally falls due to a decrease in the expiration pressure, etc. This is the above-described phrase control mechanism.

On the other hand, with respect to accent, in the case of a standard word, the accent of a word always has a remarkable pitch rise or fall from the first mora to the second mora, and
There is only one significant drop in pitch within a word. Therefore, a word consisting of n moras is (n + 1)
There are various accent types. Each accent type pays attention to the position where the pitch falls, and the 0 type, 1 type, 2 type, and 3 type (the i type is a type in which the pitch significantly decreases between the i-th and i + 1-th models) And type 0 is also called a flat accent). The rise and fall of the pitch correspond to the start point and end point of the accent command of the above-described fundamental frequency generation model.

[0013]

FIG. 6 (a) shows the shape of a phrase component by the phrase control mechanism of the conventional model, which rises relatively sharply at first, and then gradually attenuates. After the lapse of a certain time, the asymptotic reduction pattern becomes almost zero. FIG. 6A shows an example of a pattern of response speed (α = 3.0) which is generally used, but the phrase component becomes almost zero in about 2 seconds.

[0014] However, in actual speech, there are many long phrases exceeding 2 seconds. In such a case,
In the conventional phrase control mechanism, the natural fall component of the pitch in the latter half of the phrase is not expressed, and it is perceived that the pitch at the end of the phrase has risen unnaturally.

If the constant α in the above-mentioned phrase component generation function (Equation (2)) is reduced, the attenuation of the phrase component is moderated, and a natural fall component in the latter half of the phrase can be secured. (See FIG. 6 (b).) However, there is a problem that the rising speed also becomes slow at the same time, and does not match the actual rising shape of the voice. Therefore, in the conventional phrase control mechanism, when modeling the phrase component, it is necessary to compromise the values so as to appropriately satisfy both the rising response speed and the decay speed. The problem is that it is difficult to approach the shape of the froth, and furthermore, the length of the generated froth component is naturally limited.

The present invention solves the above-mentioned problems, and independently controls both the response speed and the decay speed of the rising edge so as to approximate the phrase shape of a natural voice, while providing a phrase having various lengths. An object of the present invention is to provide a fundamental frequency pattern generation method capable of adaptively generating a natural pitch pattern.

[0017]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention is to input an accent command and a phrase command for generating a prosody calculated by analyzing an input sentence, and to input a fundamental frequency on a logarithmic axis. In a fundamental frequency pattern generation method in which a pattern is represented by the sum of a phrase component corresponding to intonation and an accent component corresponding to accent, bi and ci represent the response speed of the phrase control mechanism to the i-th phrase of the input sentence. The phrase component is approximated by an impulse response function at time t represented by the following equation (4), with the determined constant ai as a constant for correcting the response strength of the phrase control mechanism to the i-th phrase of the input sentence, The method is characterized in that the constants ai, bi, and ci are set according to the length of a phrase to which a fundamental frequency is assigned.

G _pi (t) = a _i · t ^ci · exp (−b _i t) (4)

In setting the pattern shape, (c)
_i / b _i ) is preferably set in the range of 0.2 to 0.4, and b _i is 1.0 to
It is preferable to set it in the range of 4.5.

[0020]

According to the above equation (4), the response speed of the rise, that is, the time from the start of the phrase to the maximum value is given by (c _i / b _i ) seconds. On the other hand, since the decay rate is proportional to the constant bi, given frame - controls the constant b _i according to the length of the figure, by controlling the c _i in response to the rising of the response speed, frame - the's The rising response and the falling response can be controlled independently. The constant a _i can be set so that the maximum value of the function (Equation (4)) becomes a desired value, and the constants b _i and c
It acts so as to correct the response strength (maximum value of the function) of the function that changes by adjusting _i to a desired value.

FIG. 2 shows a pattern generated by the function formula (4). FIG. 2A shows an example in which the rising response speed is fast and the attenuation is slow, and FIG. The response is slow and the decay is fast. Thus, according to the present invention, it is possible to arbitrarily set the rising and falling pattern shapes by adjusting the constant.
Therefore, when modeling the phrase component, a pattern most suitable for the shape of the phrase component of natural speech can be given without being restricted by either the rising or falling response speed.

Further, according to the present invention, the constants a _i , b _i ,
c _i is frame as described above - can be set according to the length of the figure. Specifically, a plurality of sets of constants (a _i , b _i , c _i ) are prepared in advance, and the most suitable combination can be selected according to the length of the phrase component to be generated. so,
It is possible to correspond to phrases of various lengths. The above problem is solved by the above operation.

[0023]

FIG. 1 is a functional block diagram of an apparatus showing a first embodiment of the present invention. A character string input section 10, a character string analysis section 11, a phrase control mechanism 12, and an accent control mechanism 1 are shown.
3, a constant setting unit 14, a vocal cord vibration mechanism 15, and a fundamental frequency output terminal 16. Hereinafter, a method of generating a fundamental frequency in the first embodiment will be described.

First, a kana character string with a prosody symbol of a sentence to be converted into speech is input from the character string input unit 10. A kana character string with a prosody symbol is, for example, a character string shown in Table 1, and includes a kana character string corresponding to reading of a sentence to be converted into speech, a phrase symbol, an accent symbol, a pause symbol (separator symbol). And so on.

[0025]

[Table 1]

The character string analyzing unit 11 determines a phrase command which is an input of the phrase control mechanism 12 based on the input kana character string with a prosody symbol, and outputs an accent command which is an input of the accent control mechanism 13. decide.

The magnitude of the phrase command (A _{pi in} equation (1))
Is determined as shown in Table 2 according to the type of phrase symbol inserted in the input character string.

[0028]

[Table 2]

The start time of the phrase command is also determined according to the insertion position of the phrase symbol in the kana character string in the input character string. For example, in the character string shown in Table 1, the phrase symbol P1 of the first phrase is usually 100 to 100% based on the beginning of the first syllable "K" of the phrase.
It is set at the point in time when it goes back 200 ms.

The size of the accent command and the time of the command are determined according to the position and type of the pause symbol and the accent symbol. Accent marks are inserted immediately after a syllable with an accent nucleus to indicate the accent position of a word and to indicate the strength of the accent depending on the type. Table 3 shows the size of the accent command according to the type of accent symbol. Flat words (type 0) for words without accent marks
It is regarded as an accent and is given the magnitude of the command in Table 3 as well.

[0031]

[Table 3]

Although the start time of the accent command differs depending on the accent type, the first or second word of the word is used.
La vowel start time can be determined as a reference. The end time of the command is obtained based on the vowel start time of the next mora of the accent nucleus according to the position of the accent mark. In the case of the flat type, the last mora of the word is the reference.

The phrase command and the accent command determined as described above are sent to the phrase control mechanism 12 and the accent activation control mechanism 13, respectively.

The phrase control mechanism 12 calculates the expression (4), which is an impulse response function for a given phrase command, to generate a phrase component. Prior to this, the phrase to be synthesized is calculated. The constants a _i in equation (4),
The values of b _i and c _i are set by the constant raw setting unit 14.
In setting this constant value, it is necessary to previously determine a value suitable for the phrase component of the actual voice by observation, but in this embodiment, a combination of three types of constants is used in accordance with the phrase length. Shall be used.

FIG. 4 shows the relationship between the constant value and the shape of the phrase component. As shown in FIG.
The generation of a phrase component pattern according to the phrase length becomes possible.

The constant setting unit 14 selects one of the above three combinations of constants based on the number of phrases in the phrase obtained from the character string analysis unit 11 and selects the phrase control mechanism 12 Output to The length of the phrase is calculated from the duration of each phoneme composing the phrase. However, the relationship between the phrase length and the length of the phrase component is not so strict, so this embodiment In the example, the number of moras easily obtained from the input character string is substituted.

On the other hand, the accent control mechanism 13 calculates an accent component by the step response function of the equation (3) in response to the accent command given from the character string analysis unit 11.

In this way, the phrase component and the accent component for all commands are added, and the vocal cord vibration mechanism 1 is added.
5 and the vocal cord vibrating mechanism 15 adds the lower limit value ln F of the logarithmic fundamental frequency to the phrase component and the accent component.
_{The min} is added to obtain a logarithmic fundamental frequency, and exponential conversion is performed to output a fundamental frequency F0 (t) from an output terminal 16.

FIG. 3 is a block diagram showing a second embodiment of the present invention.
1, phrase control mechanism 22, phrase pattern table 2
3, an accent control mechanism 24, an accent pattern table 25, a vocal cord vibration mechanism 26, a logarithmic conversion table 27, and a fundamental frequency output terminal 28. The second
A method of generating a fundamental frequency in the embodiment will be described.

A character string input unit 2 for inputting a kana character string with a prosody symbol and generating a phrase command and an accent command
0, the operation of the character string analysis unit 21 is the same as in the first embodiment, and a description thereof will be omitted. The phrase command and the accent command determined by the character string analysis unit 21 are input to the phrase control mechanism 22 and the accent control mechanism 24, respectively.

In the phrase control mechanism 22 and the accent control mechanism 24, each component is generated based on the command. At this time, instead of calculating the above equations (3) and (4) one by one, in this embodiment, the phrase control is performed. The response patterns of the impulse response function of the pitch control mechanism 22 and the step response function of the accent control mechanism 24 are calculated in advance and stored in the phrase pattern table 23 and the accent pattern table 25, respectively. The response value of each component is calculated with reference to this. In the phrase pattern table 23, the three types of phrase patterns in the first embodiment are stored, and the phrase control mechanism 22 is stored.
Then, a suitable phrase pattern is selected according to the number of moras output from the character string analyzing unit 21.

The vocal cord vibrating mechanism 26 also requires an exponential conversion when converting the logarithmic fundamental frequency to the fundamental frequency. Therefore, a logarithmic conversion table 27 is provided here. In many cases, the pitch frequency (the reciprocal of the fundamental frequency) is actually used for speech synthesis instead of the fundamental frequency. Therefore, the logarithmic conversion table 27 is used as a table for directly converting the logarithmic fundamental frequency to the pitch period. Is also good.

[0043]

As described above in detail, according to the fundamental frequency pattern generation method of the present invention, the rise component and the decay speed of the phrase component in the fundamental frequency pattern on the logarithmic axis are separately controlled. Since it is approximated as an impulse response function that can be performed, a model closer to the natural sound phrase shape can be set without being restricted by either the rising response speed or the decay speed in modeling the phrase component. .

Further, in the present invention, the shape of the phrase component is selected according to the length of the phrase for which the fundamental frequency pattern is to be generated. A phrase component can be generated. Therefore, by performing speech synthesis using the fundamental frequency pattern generation method according to the present invention, a synthesized sound having more natural intonation can be generated.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of an apparatus showing a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of a pattern generated by Expression (4).

FIG. 3 is a functional block diagram of an apparatus showing a second embodiment of the present invention.

FIG. 4 is a diagram showing a relationship between a constant value of Expression (4) and a shape of a phrase component.

FIG. 5 is a block diagram showing a conventional fundamental frequency pattern generation model.

FIG. 6 is a diagram showing the shape of a phrase component according to a conventional method.

[Explanation of symbols]

 Reference Signs List 10 character string input section 11 character string analysis section 12 phrase control mechanism 13 accent control mechanism 14 constant setting section 15 vocal cord vibration mechanism 16 fundamental frequency output terminal 20 character string input section 21 character string analysis section 22 phrase control mechanism 23 Phrase pattern table 24 accent control mechanism 25 accent pattern table 26 vocal cord vibration mechanism 27 exponential conversion table 28 fundamental frequency output terminal

Continuation of front page (56) References JP-A-64-28695 (JP, A) JP-A-2-129700 (JP, A) JP-A-62-138898 (JP, A) Fujisaki, Sudo “Japanese word accent Model of Fundamental Frequency Pattern and Its Generation Mechanism ”, Journal of the Acoustical Society of Japan, Vol. 445-453 (1971) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A prosody calculated by analyzing an input sentence.
Enter the accent command and phrase command to generate,
The fundamental frequency pattern on the logarithmic axis
Corresponding phrase component and Accen corresponding to accent
In fundamental frequency pattern generation process represented by the sum of the bets component, deflection for the i-th phrase input sentence bi and ci
Input ai, a constant that determines the response speed of the dose control mechanism
Of the response of the phrase control mechanism to the ith phrase of
The phrase component is Gpi as a constant for correcting the strength.
(T) = time t where ai · tci · exp (−bit)
And the constant a
A basic frequency pattern generation method, wherein i, bi, and ci are set according to the length of a phrase to which a basic frequency is assigned.