JPH01219900A - Speech synthesizing device - Google Patents

Abstract

PURPOSE:To obtain a synthetic voice with superior naturalness by providing a sound source waveform generating means which generates a sound source waveform with an amplitude corresponding to the amplitude of a sound source and a sound source spectrum control means which varies the spectrum characteristics of the generated sound source waveform according to the sound source amplitude. CONSTITUTION:A sound source waveform generating means 23 generates the sound source waveform 26 with the amplitude corresponding to the sound source amplitude 25. A sound source spectrum control means 24 varies the spectrum characteristics of the sound source waveform 26 generated by the sound source waveform generating means 23 according to the sound source amplitude 25. Consequently, the spectrum characteristics of the sound source waveform 26 are varied according to the value of the sound source amplitude 24 and corresponds well to variation in actual sound source waveform, and this sound source waveform 26 is used to drive a filter for speech synthesis (voice channel filter radiation characteristic filter, etc.), thereby synthesizing a synthetic voice with superior naturalness.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a speech synthesis device that synthesizes speech waveforms by modeling the speech generation process as a linear system and driving a filter with the sound source waveform, and in particular, spectral control of the sound source waveform in the sound source section. Regarding the technology, it is possible to change the pulse shape of the voiced sound source waveform depending on the strength of the vocalization, and thereby generate a sound source waveform with an amplitude corresponding to the sound source amplitude, with the aim of synthesizing synthetic speech with excellent naturalness. and a sound source spectrum control means for varying the spectrum characteristics of the sound source waveform generated by the sound source waveform in accordance with the sound source amplitude.

[Industrial application field]

The present invention relates to a speech synthesis device that synthesizes speech waveforms by modeling the speech generation process as a linear system and driving a filter with the sound source waveform, and particularly relates to a spectrum control technique of the sound source waveform in the sound source section.

[Conventional technology]

In order to perform efficient speech synthesis, it is necessary to appropriately model the human speech production process and configure a system based on that model. A typical model of the speech generation process is a model expressed by a linear system.

The configuration of such a speech generation model is shown in FIG. As shown in the figure, the vocal tract filter 2 is
The radiation characteristic filter 3 is further driven by the output of the vocal tract filter 2 to perform speech synthesis.

The sound source unit 1 is a part that models a pulse flow of air caused by vocal fold vibration in the case of a voiced sound, and models a turbulent flow of air flowing into the vocal tract in the case of an unvoiced sound. Vocal tract filter 2
is a part that models the transfer characteristics of airflow within the vocal tract. The radiation characteristic filter 3 is a part that models the radiation characteristics of sound waves caused by airflow at the lips.

Now, in FIG. 4, the sound source spectrum at the output of the sound source section 1 is S (f), and the vocal tract transfer function at the vocal tract filter 2 is (
If the filter characteristic) is T (f) and the radiation characteristic (filter characteristic) of the radiation characteristic filter 3 is R (f),
The spectral characteristics P (f) corresponding to the change in sound pressure observed at a point separated from the lips are S (f), T (f
), R(f), and is expressed as the following equation (1).

P (f) = S (f) ・T (f) ・R(
f)... (1) In order to realize such a model, various systems have been proposed in the past. Here, a Klatt-type speech synthesizer, which is a typical formant-type speech synthesizer, will be explained as an example.

FIG. 5 shows a block configuration of the Klatt type speech synthesizer. In the figure, a sound source section 1, a vocal tract filter 2, and a radiation characteristic filter 3 correspond to those in FIG. 4, respectively. The sound source section 1 includes a voiced sound source section 4 that generates a voiced sound source waveform 9.
and an unvoiced sound source section 5 that generates an unvoiced sound source waveform 10. In addition, the vocal tract filter 2 has a voiced sound source waveform 9
It is composed of a voiced sound filter 6 driven by a voiced sound filter 6, an unvoiced sound filter 7 driven by an unvoiced sound source waveform 10, and an adder 8 that adds the outputs of each of the filters 6. Ru.

In the above configuration, the voiced sound filter 6 and the unvoiced sound filter 7 each have a second-order cyclic digital filter shown in FIG. This is realized by connecting the basic filters shown in FIG. 6 in series, and the vocal tract filter 7 for unvoiced speech is realized by connecting them in parallel.

In Figure 6, 12, 16.17 are coefficients A, B, respectively.
13 is an adder that adds the outputs of the multipliers 12, 16, and 17, and 14 and 15 are delay elements that delay the signal by a unit sampling time, respectively. Then, where T is the sampling period and n is an integer, the multiplier 12
Input x (nT) is input to , and output y from adder 13
(nT) is output, and the output y (nT) is output to the multiplier 16.
The signal y(nT-T) which is delayed by one sampling time by the delay element 14 is input, and the output y(nT) is input to the multiplier 17.
If a signal y (nT - 2T) obtained by delaying T) by two sampling times with a delay element 14.15 is input, then the input x (nT) and the output y (nT) of the basic filter in Fig. 6 are input.
The relationship T) is expressed by a difference equation such as the following equation (2).

y (nT)=/lx (nT)+B-x (nT-T
)+C-x (nT-2T)...(2) Also,
The coefficients A, B, and C in the multipliers 12, 16, and 17 of the basic filter in FIG. 6 are the resonance frequency F and its bandwidth BW.
is calculated from the following equations (3) to (5). Here, T is the sampling period, π is pi, exp is an exponential function operation, and cos is a cosine function operation.

C=-exp (-2π・BW-T)
...13) B=2・exp (−π・BW・T)
・cos (2πFT) ・ ・ ・ ・ (4) A=1−C−B ・ ・ ・ ・ ・(5) From these relationships, a plurality of series-connected basics forming the voiced sound filter 6 in FIG. Voiced sounds are synthesized by providing filters with separate resonant frequencies F and their bandwidths BW, each of which varies over time. Unvoiced sound filter 7
The same applies to

Next, the unvoiced sound source section 5 in FIG. 5 is composed of a white noise generating section that generates white noise as the unvoiced sound source waveform 10, a multiplier that controls the amplitude value, etc., although not particularly shown.

On the other hand, FIG. 7 shows a conventional general configuration of the voiced sound source section 4 shown in FIG. As shown in the figure, the impulse train waveform 21
an impulse generating section 18 that generates an impulse train waveform 21, a low-pass filter 19 that receives an impulse train waveform 21 as input, and a voiced sound source waveform 9 (see FIG. 5) that multiplies its output by a sound source amplitude 22.
It consists of a sound source amplitude multiplier 20 that outputs.

Here, the impulse train generation unit 1 generates an impulse train waveform 21 at a period corresponding to a fundamental frequency FO called a pitch corresponding to the vibration of the vocal cords.

The spectrum of this waveform 21 has a comb-shaped harmonic structure having frequency components (harmonic components) of the same amplitude at the fundamental frequency FO and each integral multiple thereof.

Next, the low-pass filter 19 filters the impulse train waveform 21 having the above-mentioned spectral characteristics, thereby forming a harmonic structure having an envelope in which the amplitude gradually decreases from the fundamental frequency toward higher harmonic components. Convert to a waveform with spectral characteristics. This waveform has spectral characteristics close to the actual sound source waveform.

After the above processing, the sound source amplitude multiplier 20
2 is set, the voiced sound source waveform 9 is output, and the voiced sound filter 6 shown in FIG. 5 is driven.

The low-pass filter 19 in FIG. 7 uses one of the basic filters in FIG. 6 used for the vocal tract filter 2 in FIG.
) tr - By setting it to an appropriate value, low-pass characteristics are realized. In addition, an anti-resonance filter is actually connected to shape the shape of the spectrum of the sound source, but
It is omitted here.

[Problem to be solved by the invention]

With the conventional voiced sound source section 4 shown in FIGS. 5 and 7, a pulse train having a constant pulse shape and a period corresponding to the fundamental frequency FO is obtained as the voiced sound source waveform 9. As described above, its spectral characteristics have a harmonic structure having a constant envelope in which the amplitude gradually decreases from the fundamental frequency toward higher frequency harmonic components. Note that each frequency component changes over time according to the pitch.

However, in the voiced sound source waveform generated by the human vocal cords, even if the pitch is constant, that is, the fundamental frequency FO is constant and the pulse interval is constant, the shape of each pulse depends on the strength of the vocalization (see Figure 8 (Ta)). Or change as in (b). That is, when the vocalization is strong, it is approximated by a highly impulsive asymmetric triangular wave as shown in FIG. 8(a), and when the vocalization is weak, it is approximated by a sine wave as shown in FIG. 8(b).

However, in the conventional voiced sound source waveform generation method, although the sound source amplitude changes depending on the strength of generation, the pulse shape of the obtained voiced sound source waveform 9 (see FIGS. 5 and 7) is constant. This method has a problem in that it is not possible to differentiate the sound quality of synthesized speech depending on the strength of the utterance, and it is not possible to synthesize synthesized speech with excellent naturalness.

An object of the present invention is to make it possible to change the pulse shape of a voiced sound source waveform depending on the strength of vocalization, thereby synthesizing synthetic speech with excellent naturalness.

[Means to solve the problem]

FIG. 1 is a block diagram of the present invention. The sound source waveform generating means 23 generates a sound source waveform 26 having an amplitude corresponding to the sound source amplitude 25. The means 23 includes, for example, an impulse train generation section that generates a digital impulse train waveform having a pitch period, a digital low-pass filter that receives the impulse train waveform as input, and a digital value of the sound source amplitude 25 that is multiplied by the output of the filter. and a multiplier that outputs a voiced sound source waveform.

The sound source spectrum control means 24 includes the sound source waveform generation means 23
The spectral characteristics of the sound source waveform 26 generated by the sound source are varied according to the sound source amplitude 25. The means 24 is constituted by, for example, means for varying the filter coefficients of the digital low-pass filter in the sound source waveform generating means 23 so that the bandwidth thereof changes in direct proportion to the sound source amplitude 25.

[Function] With the above means, the spectral characteristics of the sound source waveform 26 change according to the strength of the sound source amplitude 25, and correspond well to changes in the actual sound source waveform. By driving synthesis filters (vocal tract filter, radiation characteristic filter, etc.), a synthesized speech with excellent naturalness is synthesized.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail.

First, the overall configuration of the speech synthesis device targeted by the present invention is as follows.
Since it is the same as the Klatt type speech synthesizer shown in FIG. 5, which has already been explained, its explanation will be omitted.

FIG. 2 is a block diagram of a first embodiment of the present invention relating to the voiced sound source section 4 of FIG. First, the impulse generation section 23 generates an impulse train waveform 27 at a period corresponding to the fundamental frequency FO.

Next, this waveform 27 is filtered by a low-pass filter 24 whose spectral characteristics are controlled by a filter coefficient controller 25. The low-pass filter 24 is composed of one basic filter, which is the second-order cyclic digital filter shown in FIG. 7, which has already been explained.

The filter coefficient control unit 25 calculates the coefficients of the low-pass filter 24 according to the sound source amplitude 28.

The sound source amplitude multiplier 26 multiplies the output of the low-pass filter 24 by the sound source amplitude 28 (digital value) to obtain
A voiced sound source waveform 9 (see FIG. 5) is output.

Next, the operation of the first embodiment having the above configuration will be described below.

First, the impulse train waveform 27 generated by the impulse train generating section 23 is a waveform in which impulses are arranged at a period corresponding to the fundamental frequency FO (pitch) corresponding to the vibration of the vocal cords, and its spectral characteristics are A comb-shaped harmonic structure has frequency components (harmonic components) of the same amplitude at each frequency that is an integer multiple of .

Next, the low-pass filter 24 filters the impulse train waveform 27 to convert it into a voiced sound source waveform 9 as shown in FIG. When is small, Figure 8 ('b)
It is converted into a voiced sound source waveform 9 as shown in FIG.

Here, each voiced sound source waveform in FIGS. 8(a) and (b) is
As already mentioned, it shows a waveform that is close to the waveform actually generated by humans. When a strong voice is uttered, it is approximated by an asymmetric triangular wave with strong impulsiveness, as shown in +8) in the same figure. This means that the spectral characteristics of the voiced sound source waveform have a gentle envelope characteristic in which the amplitude does not decrease much from the fundamental frequency to the high-frequency harmonic components, resulting in a harmonic structure with many harmonic components. It is shown that.

On the other hand, when vocalization is weak, the sound becomes sinusoidal as shown in FIG. 2(b). This means that the spectral characteristics of the voiced sound source waveform have a harmonic structure with an envelope characteristic in which the amplitude rapidly decreases from the fundamental frequency toward higher frequency harmonic components.
indicates that it consists mostly of the fundamental frequency component.

Based on the above facts, the transfer characteristics of the low-pass filter 24 in FIG. Sometimes, a characteristic is required in which the cutoff frequency is so low that high frequency harmonic components are not passed.

In order to achieve the above characteristics when the basic filter shown in FIG. 6 is used as the low-pass filter 24 shown in FIG. When it's small, it's fine if it gets narrower.

Therefore, the filter coefficient control unit 25 in FIG. 2 fixes the resonant frequency F to 0, and changes its bandwidth BW so that the following equation is satisfied: (6), and then, The values of the coefficients A, B, and C of the basic filter shown in FIG. Note that AV is the value of the sound source amplitude 28, α and β are coefficients determined experimentally, and α, β>0.

Through the above operations, the voiced sound source waveform 9 obtained in the first embodiment shown in FIG.
A waveform as shown in FIG. 1 or (b) is obtained, which corresponds well to the waveform of an actual human voiced sound source. By using such a voiced sound source waveform 9 and performing speech synthesis using the Klatt type speech synthesizer shown in FIG. 5, it is possible to make the synthesized speech particularly natural in the voiced sound section.

Next, FIG. 3 is a block diagram of a second embodiment of the present invention relating to the voiced sound source section 4 of FIG. 5. In this embodiment, another low-pass filter 29 is provided between the impulse generator 23 and the low-pass filter 24 of the first embodiment shown in FIG.
It is configured with the following.

In the above configuration, the filter coefficients of the low-pass filter 29 are set so that the transmission characteristic of the low-pass filter 29 generates a sound source spectrum corresponding to the case where the sound source amplitude 28 is the strongest. The transfer characteristic of the low-pass filter 24 is controlled by the filter coefficient control section 25 in accordance with the magnitude of the sound source amplitude 28 in the same manner as in the first embodiment. For example, if the low-pass filters 29 and 24 in FIG. 3 are each configured with one basic filter in FIG. The BW is set wide to generate a spectrum corresponding to strong vocalizations, and the bandwidth BW at the resonance frequency of the low-pass filter 24 is controlled to be narrow in the case of strong vocalizations and narrow in the case of weak vocalizations.

With the above configuration, it is possible to perform more fine-grained spectrum control than in the first embodiment shown in FIG.

In the embodiment of the present invention shown above, for example, if a filter with a higher order than the one in FIG. 6 is used as the low-pass filter 24 in FIG. Fine-grained spectrum control can be achieved with a single low-pass filter without having to connect two filters.

Furthermore, the present invention is not limited to the low-pass filter as shown in FIG. 2 or 3; a filter whose anti-resonance characteristic changes depending on the strength of vocalization may be added.

Furthermore, the present invention is not limited to Klatt-type speech synthesizers, but can be broadly applied to speech synthesizers that model the speech generation process using a linear system in which a filter is driven by a sound source waveform.

〔Effect of the invention〕

According to the present invention, it is possible to vary the spectrum of a sound source waveform according to the sound source amplitude, and by using this to drive a filter for speech synthesis, it is possible to synthesize synthesized speech with excellent naturalness. It becomes possible.

In particular, by controlling the impulse train waveform to pass higher harmonic components in direct proportion to the sound source amplitude, it is possible to generate a voiced sound source waveform that corresponds well to human speech.

At this time, the impulse train waveform has a resonance frequency of 011z
By filtering with a low-pass filter whose bandwidth changes in direct proportion to the sound source amplitude, a voiced sound source waveform of good quality can be obtained.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the present invention, FIG. 2 is a configuration diagram of the first embodiment of the present invention, and FIG. 3 is a block diagram of the present invention.
A block diagram of the second embodiment of the present invention, FIG. 4 is a block diagram of a speech generation model, FIG. 5 is a block diagram of a Klatt-type speech synthesizer, and FIG. 6 is a basic diagram of a Klatt-type speech synthesizer. FIG. 7 is a configuration diagram of a conventional filter, and FIGS. 8(a) and 8(fb) are voiced sound source waveform diagrams. 23... Sound source waveform generation means, 24... Sound source spectrum control means, 25... Sound source amplitude, 26... Sound source waveform. Patent Applicant: Fujitsu Limited September 1999 Blockchain Figure 1 Figure 11 of the invention's first breakthrough Figure 2 Figure 6 Base filter of Klatt type self-voice @growth device Figure 6 Conventional filter To 10 years old, Figure 7. Teku, multicolor, one time (a) width → time shoulder voice, figure 8.

Claims (1)

[Scope of Claims] 1) In a speech synthesis device that synthesizes a speech waveform by driving a filter with a sound source waveform, a sound source waveform generating means ( 23) and the sound source waveform (2
6) Sound source spectrum control means (24) for varying the spectrum characteristics according to the sound source amplitude (25). 2) The speech synthesis is performed by digital signal processing, the sound source waveform (26) is a digital voiced sound source waveform, and the sound source waveform generating means (23) is an impulse generator that generates a digital impulse train waveform with a pitch period. Department and
The sound source spectrum control means is composed of a digital low-pass filter that receives the impulse train waveform as input, and a multiplier that multiplies the output of the filter by a digital value of the sound source amplitude (25) to output the voiced sound source waveform. (24) is the sound source amplitude (
By varying the filter coefficient of the digital low-pass filter according to (25), when the sound source amplitude (25) is small, among the fundamental frequency component and its harmonic component corresponding to the pitch period in the impulse train waveform, 2. The sound source according to claim 1, wherein the control is performed so that only the vicinity of the fundamental frequency component is passed, and when the sound source amplitude (25) is large, the control is performed so that high frequency harmonic components are also passed. Speech synthesizer. 3) The digital low-pass filter has a resonant frequency of 0.
Hz transmission characteristic, and the sound source spectrum control means (
3. The speech synthesis apparatus according to claim 2, wherein the filter coefficient of the digital low-pass filter is varied so that the bandwidth of the digital low-pass filter changes in direct proportion to the sound source amplitude.