JPH01237700A - Basic frequency pattern forming device - Google Patents

Abstract

PURPOSE:To eliminate the influence by the fluctuation in pronounced speeches by providing a parameter forming means which can control the rise and fall of a basic frequency pattern and the amplitude thereof and a basic frequency forming means which forms the basic frequency signals in the vowel central part of respective moras in accordance with a prescribed mathematical expression. CONSTITUTION:A parameter control section 11 forms the parameter signal which can control the rise and fall of the basic frequency pattern and the amplitude thereof. A basic frequency calculating section 12 receives the parameter signal and forms the basic frequency signals of the vowel central parts for each of the moras in accordance with the prescribed mathematical expression. Since the basic frequency signals are time-normalized with the respective moras and are, therefore, not affected by the fluctuation in the pronounced speeches. A basic frequency interpolation section 13 receives the discrete basic frequency signals and makes interpolation between the respective basic frequency signals. The frequency pattern forming device which can control the rise and fall of the basic frequency pattern and the amplitude thereof and is not affected by the fluctuation in the pronounced speeches is thereby obtd.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a fundamental frequency pattern generation device used for speech synthesis, and particularly to a fundamental frequency pattern generation device for improving the naturalness of synthesized speech.

[Prior Art and Problems to be Solved by the Invention] Conventionally, a point pitch pattern model has been cited as a typical model for describing the fundamental frequency pattern in normal Japanese speech. This is a discrete model limited only to the fundamental frequency value of each vowel center.

FIG. 2 is a graph showing, by way of example, a dot pitch pattern showing a word with a 5-mora type 3 accent. This graph, for example, was published in the November 1973 IEICE Transactions Special Issue (Trans, IECE '7B/11 V
ol, 56-D No.

11).

Referring to FIG. 2, 1 to 5 on the horizontal axis indicate the position of each mora of this word, and the vertical axis indicates frequency (Hz). Each point (101) to 105) indicates the value of the fundamental frequency at each vowel center of each node.

Such a dot pitch pattern exhibits a substantially constant shape depending on the mora length and accent type, regardless of the content of the word. That is, words with a 5-mora type 3 accent generally have a dot-pitch conflict pattern as shown in FIG. In this model, only 5 fundamental frequency values (101) to 105) are obtained at the center of each vowel, but by linearly interpolating these values, it is possible to obtain good synthesized speech using this model. Can be done.

Since the fundamental frequency pattern differs depending on the mora length and accent type, various fundamental frequency patterns corresponding to each are required, and modeling for this purpose has also been proposed.

However, such conventional models are for normally uttered voices, and cannot control the amplitude of the fundamental frequency pattern.

In fact, it is known that the dot pitch pattern is hardly affected by variations in vocalization rate, but is significantly affected not only by the pitch of the vocalization but also by the loudness of the vocalization. Therefore, in order to synthesize more natural speech, a model is desired that can uniformly and mathematically control the rise, fall, amplitude, etc. of the fundamental frequency pattern.

An example of a model for mathematically controlling such a fundamental frequency pattern is a critical damping second-order system model. This is a continuous model that generates the entire pattern based on specifications such as the accent command time point and amplitude size. Thereby, the fundamental frequency pattern can be approximated with high accuracy, and quantitative handling using parameters becomes easier.

The approximation formula can be found, for example, in the Journal of the Acoustical Society of Japan (Vol. 27, No. 9, pp. 445-453, 1971) by Fujisaki et al. entitled "Model of fundamental frequency patterns of Japanese word accents and their generation mechanism". It is shown below.

PO(t) −FIIin-exp[Gv(t-to)
-Gv(t-t3)+Ga(t-tl)-Ga(t-t
2)-g2] ・”Formula (1) Gy(t)
−Avαt −exp (−αt) −Formula (2
)Ga(i) -A a[1-(1+βt) ・exp
(-β1)] - Equation (3) where FO(t) is the logarithmic fundamental frequency pattern, Fm1n is the lowest frequency that can oscillate, time to and t3 are the start and end points of the tone command, respectively, and the time t1 and t2 are the start and end points of the accent command, respectively, and g2 is the value of the input signal corresponding to stopping oscillation. Further, Av is an amplitude parameter indicating the magnitude of the response of the tone component, Aa is an amplitude parameter indicating the magnitude of the response of the accent component, and α and β are time components indicating the response speed of each component.

FIG. 3 is a block diagram showing a fundamental frequency control mechanism for obtaining a fundamental frequency pattern using equations (1) to 3). This block diagram is described in the aforementioned document.

Referring to FIG. 3, the tone component generator 6 generates a tone command signal S.
9 and outputs tone component Gv (t). The speech tone command signal S9 is a unit pulse that rises at time to and falls at time t3. On the other hand, the accent component generating section 7 receives the accent command signal SIO and outputs the accent component Ga (t). The accent command signal SIO is a unit pulse that rises at time t1 and falls at time t2. (Here, there is a relationship of to<tl<t2<t3.) Both tone component generating section 6 and accent component generating section 7 operate according to the above-mentioned equations (1) to 3).

The fundamental frequency pattern synthesis unit 8 adds the tone component Gv(t) and the accent component Ga(1) of the fundamental frequency,
A fundamental frequency pattern FO(t), which is a continuous model, is synthesized.

In order to determine each parameter of the above equations (1) to 3), Analysis by 5ynthes
A method called the Ls method (hereinafter abbreviated as AbS method) is used. This is done by giving appropriate initial values to each of the parameters mentioned above and repeatedly comparing the fundamental frequency pattern obtained by gradually changing these parameters with the actually observed fundamental frequency pattern.
This is to estimate the optimal parameters.

In this way, by the method shown in FIG.
A highly accurate fundamental frequency pattern can be obtained.

However, since the target is a continuous pattern, Ab
The problem is that iterative calculations of the S-method take a very long time, and the local fluctuations in the fundamental frequency that are actually observed (for example, when a word contains unvoiced sounds)
There is a problem in that this has a negative effect on parameter estimation.

Furthermore, conventionally, the timing of accent commands and the like is set to be a certain period of time after the start point of a vowel, so there is a problem in that overall or partial variations in speech rate have a negative effect on parameter estimation.

This invention was made to solve the above-mentioned problems, and it is possible to uniformly and precisely control the rise and fall of the fundamental frequency pattern as well as the amplitude, and it is possible to control the rise and fall of the fundamental frequency pattern as well as the amplitude with high precision. The purpose of this invention is to obtain a frequency pattern generation device that is not affected.

[Means for Solving the Problems] A fundamental frequency pattern generation device according to the present invention includes a parameter generation device that generates a parameter signal for controlling the rise and fall and amplitude of a fundamental frequency pattern, and a parameter generation device that receives the parameter signal and generates a predetermined a fundamental frequency generating means for generating a fundamental frequency signal at the center of a vowel that is time-normalized for each mora based on a mathematical formula; and an interpolating means for receiving the fundamental frequency signal and interpolating between the fundamental frequency signals. .

[Operation] In the basic frequency pattern generation device according to the present invention, the parameter generation means generates a parameter signal that can control the rise and fall and amplitude of the basic frequency pattern. The fundamental frequency generating means receives this parameter signal and generates a fundamental frequency signal of the vowel center for each mora based on a predetermined mathematical formula. Since this fundamental frequency signal is time-normalized for each mora, it is not affected by fluctuations in the uttered voice. The interpolation means receives the discrete fundamental frequency signals and interpolates between the respective fundamental frequency signals.

[Embodiment of the Invention] FIG. 1 is a block diagram schematically showing a fundamental frequency pattern generation device and a speech synthesis device using the same, showing an embodiment of the present invention.

Referring to FIG. 1, the basic frequency pattern generation device 1 includes:
a parameter control unit 11 that receives a basic frequency pattern control information signal S1 and outputs a parameter set signal S2;
A fundamental frequency calculation unit 12 receives the signal S2 and outputs a point pitch pattern signal S3 of the fundamental frequency value of the vowel center for each mora, and receives the signal S3 and interpolates it to output a fundamental frequency pattern signal S4. A basic frequency pattern interpolation unit 13 is included.

This speech synthesis device receives the fundamental frequency pattern signal S4 from the fundamental frequency pattern generation device 1, and the signals S5, S6, and S7 from the power pattern generation section 2, phoneme duration determination section 3, and spectral pattern generation section 4, respectively.
It includes a synthesized speech generating section 5 connected to receive the synthesized speech. A synthesized speech signal S8 is output from the synthesized speech generation section 5.

Next, the operation of the basic frequency pattern generation device 1 will be explained.

The parameter control unit 11 receives a control information signal S1 from the outside. The control information signal S1 includes, for example, the mora, the number of syllables, the average frequency, etc. of the word for which the fundamental frequency pattern is generated. The parameter control unit 11 uses these control information to output a parameter set signal S2 for generating a fundamental frequency pattern. This parameter set signal S2 includes parameters that can control the rise and fall as well as the amplitude of the fundamental frequency pattern.

Various mathematical formulas can be used to find the fundamental frequency value, such as one using the response of the critical damping quadratic system mentioned above, or one using polynomial expansion. Hereinafter, a method using the response of the critical braking secondary system will be explained.

This is based on the above-mentioned formulas (1) to 3). For example, in the case of a word with a 5-mora 3-type accent,
The end command of the accent component is assumed to be t2-3 (mora). Generally, for words with n-mora i-type accent, t
2= i (mora) (however, when i-〇, t2-n
), it is possible to correspond to various accent types. Other parameters are similarly determined.

The fundamental frequency calculating section 12 receives the parameter set signal S2, calculates the fundamental frequency value at the center of the vowel for each mora, and generates a point pitch pattern.

The basic frequency pattern interpolation unit 13 receives the point pitch pattern signal S3, performs linear interpolation between the point pitch patterns, determines the shape of the entire basic frequency pattern, and outputs the basic frequency pattern signal S4.

In addition to the fundamental frequency pattern signal S4, the synthesized speech generation section 5 receives a power pattern signal S5 from the power pattern generation section 2 and a phoneme duration signal S6 from the phoneme duration determination section 3.
, and the spectral pattern signal S7 from the spectral pattern generation section 4, generates a synthesized sound, and outputs a synthesized speech signal S8.

These power pattern signal S5, phoneme duration signal S6
There are many conventional methods for generating or synthesizing the spectral pattern signal S7.
Either can be used.

Note that the above description of the mathematical formulas used in the fundamental frequency calculation section 12, the linear interpolation performed in the fundamental frequency pattern interpolation section 13, etc. is merely an example, and the scope of the present invention is not limited thereby. It's not something you can do.

Therefore, according to the present invention, since the fundamental frequency pattern obtained by the fundamental frequency of the vowel center part in each mora is used, it is possible to obtain stable control parameters with local fluctuations removed. Since the time axis is normalized, the influence of variations in speaking speed can be removed. In addition, it is controlled by a mathematical formula that includes parameters that can control the rise and fall of the fundamental frequency pattern as well as the amplitude, so it is possible to use highly accurate quantitative expressions that can accommodate not only mora length and accent type, but also a variety of tones. It becomes possible to perform control and obtain more natural synthesized speech.

Furthermore, since the amount of data handled when determining control parameters is very small compared to conventional continuous models, these parameters can be estimated in a short time. It is also possible to simultaneously analyze a plurality of fundamental frequency patterns using the AbS method, thereby making it easy to remove variations in the fundamental frequency patterns from voice to voice or from word to word.

[Effects of the Invention] As described above, according to the present invention, there is provided a parameter generation means that can control the rise and fall as well as the amplitude of a fundamental frequency pattern, and a fundamental frequency signal of the vowel center of each mora based on a predetermined mathematical formula. Since the basic frequency pattern includes a fundamental frequency generating means that generates a fundamental frequency pattern, it is possible to uniformly and highly accurately control the rise and fall of the fundamental frequency pattern as well as the amplitude, and it is not affected by fluctuations in vocalizations during analysis. A frequency pattern generation device is provided.

[Brief explanation of the drawing]

FIG. 1 is a block diagram schematically showing a fundamental frequency pattern generation device and a speech synthesis device using the same, showing an embodiment of the present invention. Figure 2 shows the conventional point pitch
It is a graph showing a pattern model. FIG. 3 is a block diagram showing a conventional fundamental frequency control mechanism. In the figure, 1 is a fundamental frequency pattern generation device, 2 is a power pattern generation section, 3 is a phoneme duration determination section, 4 is a spectral pattern generation section, 5 is a synthesized speech generation section, 6 is a tone component generation section, and 7 is a speech tone component generation section. Accent component generation part, 8 is synthesis part,
1.1 is a parameter control section, 12 is a fundamental frequency calculation section,
Reference numeral 13 indicates a basic frequency pattern interpolation section. Patent applicant A.T.R. Co., Ltd. Figure 2 + 2 3 4 5 (Mo2i fr, y,)

Claims (1)

[Scope of Claims] A fundamental frequency pattern generation device for generating a fundamental frequency pattern for speech synthesis, comprising parameter generation means for generating a parameter signal for controlling the rise and fall and amplitude of the fundamental frequency pattern; Discrete fundamental frequency generating means receives a parameter signal and generates a discrete fundamental frequency signal at the center of a vowel that is time-normalized for each mora based on a predetermined mathematical formula; , and interpolation means for interpolating between each fundamental frequency signal.