JP3133427B2 - Speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for synthesizing speech by superimposing information corresponding to speech units.

2. Description of the Related Art A speech synthesizer as disclosed in Japanese Patent Application Laid-Open No. 1-239292 has been known. In this speech synthesizer, the zero-phase impulse response waveform is used as a speech segment waveform, and the amplitude of the speech segment waveform and the superimposition period are randomly specified and given by random numbers, thereby giving
It is intended to synthesize natural unvoiced sounds close to human voice.

More specifically, the amplitude of the impulse response waveform is changed randomly by multiplying the impulse response waveform of the noise signal by a random value at regular intervals (approximately 0.17 msec). about,
As shown in FIG. 11, the next superimposition timing is given by a random number (an integer value from “1” to “5”) from the previous timing, whereby the superimposition is performed as indicated by R1, R2, and R3 in the figure. The matching cycle was given at random.

[0004]

However, in the above-described conventional speech synthesizer, the amplitude is made random by multiplying the impulse response waveform, that is, the speech unit waveform by a random value, and thus the synthesized speech is synthesized. There is a problem that the spectral characteristics of the unvoiced sound are lost.
That is, by making the amplitude of the speech unit waveform random, the spectral characteristics of unvoiced sound actually uttered by a human cannot be satisfactorily approximated.

In the above-mentioned conventional speech synthesizer,
Although the phase characteristics are randomized by making the superimposition cycle random, in this superimposition, since the next superimposition timing is affected by the immediately preceding timing, the overlap is not uniform, There is a problem that the power of the waveform after the superposition greatly varies, and there is a problem that the randomness of the phase characteristics is not sufficient. Therefore, this speech synthesizer has a limit in synthesizing a more natural unvoiced sound close to the real human voice.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech synthesizer capable of generating a natural synthesized speech closer to the real human voice even when synthesizing an unvoiced sound.

[0007]

In order to achieve the above object, the invention according to claim 1 comprises a randomizing means for randomizing the phase of a speech unit waveform to generate a random-phase speech unit waveform; And a waveform superimposing means for adding or superimposing the random phase speech unit waveforms generated by the random phase converting means while shifting them to synthesize an unvoiced speech waveform.

According to a second aspect of the present invention, the waveform superimposing means includes a random signal generating means for designating a superimposition timing of a random phase speech unit waveform by a random value;
And superimposing means for synthesizing an unvoiced sound waveform by shifting or adding or superimposing a random phase speech unit waveform by the superimposition timing instructed by the random signal generating means. I have.

Further, the invention according to claim 3 is characterized in that a windowing processing means for performing windowing processing on a speech unit waveform after random phase conversion is further provided.

[0010]

[0011]

According to the first aspect of the invention, when performing voice synthesis of an unvoiced sound, the phase of the voice unit waveform is randomized instead of the amplitude. As a result, the synthesized voice of the unvoiced sound obtained is sufficiently random, such as white noise, only in phase characteristics, while maintaining the spectral characteristics of unvoiced sounds actually uttered by humans.

According to the second aspect of the present invention, not only the phase of the speech unit waveform but also the timing of superposition are made random, so that the phase characteristics are further randomized.

According to the third aspect of the present invention, the window processing is performed on the speech unit waveform after the random phase conversion,
The random-phase speech unit waveforms in which the discontinuity at the beginning and end are alleviated can be superimposed while being shifted at a predetermined timing.

[0014]

[0015]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a speech synthesizer according to the present invention. Referring to FIG. 1, the speech synthesizer includes a random phase unit 2 for randomizing the phase of a speech unit waveform 1, a switching unit 3 for switching whether to generate unvoiced or voiced synthesized speech, and a voiceless unit. When the synthesized speech is generated, the speech unit waveform randomized by the random phase unit 2 is added or superimposed while being shifted, and when the voiced synthesized speech is generated, the speech unit waveform 1 is added. Addition or superimposition while shifting sequentially at a predetermined pitch cycle,
A waveform superimposing unit 4 for generating a synthesized voice.

The speech unit waveform 1 is, for example, LPC, L
It is created by performing an inverse Fourier transform on the spectral envelope of the voice obtained by a voice analysis technique such as SP or PSE, and thus represents a waveform equivalent to the inverse Fourier transformed waveform of the spectral envelope of the voice. .

The randomizing section 2 randomizes only the phase of the speech unit waveform 1 while maintaining the spectral envelope characteristic of the given speech unit waveform 1 as it is. . FIG. 2 is a diagram showing an example of the configuration of the random phase shifter 2. In the example of FIG. 2, the random phase shifter 2 performs a Fourier transform on the speech unit waveform 1 to obtain a spectrum S, A random number generator 6 for generating a random number R, and an inverse Fourier transform unit 7 for randomly giving the phase of the spectrum S obtained by the Fourier transform unit 5 with the random number R, and re-forming the time domain waveform by inverse Fourier transform. Have been.

The wavelength superimposing unit 4 is configured to add or superimpose a random-phased speech unit waveform while shifting it at a predetermined pitch cycle when generating a synthetic voice of unvoiced sound. Can be.

Next, the operation of the speech synthesizer having such a configuration will be described. When attempting to generate an unvoiced synthesized voice, the switching unit 3 switches to the “unvoiced” side. Note that such a switching operation may be performed, for example, by adding information indicating whether a frame is a voiced section or an unvoiced section to the switching unit 3, or by setting the positive and negative values of the pitch data respectively. It may be performed by adding to the switching unit 3 as voiced sound information and unvoiced sound information. When the random phase conversion unit 2 has a configuration as shown in FIG. 2, for example, the spectrum S is obtained by Fourier-transforming the speech unit waveform 1 as shown in FIG. Next, the phase of the spectrum S is randomly given by a random number R, and the waveform in the time domain is again formed by the inverse Fourier transform. The waveform thus obtained has a randomized phase as shown in FIG. 3 (b), and the waveform superimposing unit 4 uses the randomized phase in FIG. 3 (b).
Are added or superimposed while being shifted at a predetermined pitch period F, for example, as shown in FIG.
Thereby, a synthesized voice of unvoiced sound can be obtained.

As described above, in this embodiment, at the time of unvoiced synthesized speech, the phase of the speech unit waveform 1 is basically randomized instead of the amplitude. Regarding the spectral characteristics, it is possible to maintain the spectral characteristics of the unvoiced sound actually uttered by a human and make the phase characteristics sufficiently random like white noise only for the phase characteristics. As a result, it is possible to generate a more natural unvoiced synthesized voice that is closer to the real human voice than in the past.

In the above-mentioned example, the randomization of the phase is performed in real time by activating the random phase shifter 2 at the time of speech synthesis. This may be stored in a memory or the like in advance, and a random phase speech unit waveform stored in the memory or the like may be read out and sent to the waveform superimposing unit 4 at the time of voice synthesis of an unvoiced sound.

In the above example, the waveform superposition unit 4 converts the random phase speech unit waveform into the waveforms shown in FIGS.
As shown in (c), the layers are overlapped while being shifted at a predetermined pitch cycle. In this case as well, the randomness of the phase characteristics can be improved to a better level as compared with the related art.
In addition to randomizing the phase of the speech unit waveform, the superposition timing of the speech unit waveform is also randomized, so that the spectral characteristics of unvoiced sound as synthesized speech While the spectral characteristics of No. 1 can be faithfully reproduced and maintained,
As regards the phase characteristic, it can be made more random like almost perfect white noise.

FIG. 4 is a diagram showing an example of the configuration of the waveform superposition unit 4 intended to make the timing of superimposing the speech unit waveform 1 random.

In the configuration example of FIG. 4, the waveform superimposing section 4 includes a random signal generating section 11 for generating a random signal and a pitch cycle generating section 12 for generating a pulse having a predetermined pitch cycle.
When the voice synthesis of the unvoiced sound is performed, the random signal generation unit 11
And a switching unit 13 that switches to select a random signal from the pitch period generating unit 12 during voice synthesis of a voiced sound.
And a superposition unit 14 for superposing the speech unit waveforms from the switching unit 3 as shown in FIG.

When the wavelength superimposing section 4 is configured as shown in FIG. 4, when synthesizing unvoiced sound, the switching section 13 applies a random signal (for example, a random pulse) from the random signal generating section 11 to the superposing section 14. This is given as the timing of superposition. At this time, when the random phase speech unit waveform from the switching unit 3 is sent to the wavelength superposition unit 4, the superimposition unit 14 of the wavelength superposition unit 4 converts the random phase speech unit waveform with the random signal from the switching unit 13. The random phase speech unit waveforms are superimposed while being shifted at the designated superimposition timing, and an unvoiced speech waveform is synthesized. In this way, the synthesized voice of the unvoiced sound obtained has not only the phase of the speech unit waveform but also the timing of superposition being random, so that the phase characteristics are further randomized and more similar to human voice. Natural synthesized speech.

In the configuration example shown in FIG. 4, a windowing processing unit 15 may be provided at a stage preceding the overlapping unit 14 as shown. The windowing processing unit 15 performs windowing processing such as a Hamming window and a Hanning window on the random phase speech unit waveform to reduce discontinuity at the start and end thereof. When a window WIN as shown in FIG. 5A is set in such a windowing processing unit 15, a random-phase speech unit waveform as shown in FIG. Due to the window WIN of the processing unit 15, the waveform is deformed as shown in FIG. 5C, and the discontinuity at the start and end of the waveform in FIG. 5B can be reduced. Are superimposed while shifting at predetermined timing the random-phase speech unit waveforms in which the discontinuity is alleviated. As a result, it is possible to generate a natural synthesized speech that is closer to the real human voice than the discontinuity is reduced.

Further, in the configuration of FIG. 4, the random signal generator 11 can be configured as shown in FIG. That is, in the configuration example of FIG. 6, the random signal generator 11, a random number, i.e. the random value r _n (eg "-2
0 a random number generator 51 for generating integer values) between "from" 20 ", the current random value r _n and the random value r _n-1 and the difference at the time of the previous (r _n -r _{n -1} )
When, and a differential value (r _n -r _n-1) and adds the adder 53 from the differentiator 52 and a fixed period T (for example, "50").

When the random signal generator 11 has such a configuration, the timing period l _n (n = 1, 2, 3,...) Of superposition of random phase speech unit waveforms
As shown in FIG. 7, it is obtained by adding a fixed period T the difference value _{(r n -r n-1)} , variations in the predetermined cycle T is given to the superposition section 14 such that the random be able to. That is, conventionally, as shown in FIG. 11, the superimposition timing is given by a random number from the previous timing, but the random signal generation unit 11 of FIG.
A fixed period T
Is given to the superimposing unit 14 so that the variation from the randomization is random, so that the overlap of the speech unit waveforms becomes uniform and the phase characteristics become sufficiently random, so that the human A natural synthesized voice similar to the real voice of the subject can be generated.

As described above, the speech synthesizer is configured as shown in FIGS. 1, 4 and 6, and the phase of the speech unit waveform is made random. In addition, the superposition timing is also made random. Is effective in synthesizing an unvoiced sound that is closer to an unvoiced sound actually generated by a human. However, various modifications can be made based on the basic configuration of FIG. For example, instead of making the timing of superposition random, it is also possible to make the amplitude of the random phase speech unit waveform random.

FIG. 8 and FIG. 10 are diagrams showing such a configuration example. In the configuration example of FIG. 8, the wavelength superposition unit 4 includes a period generation unit 21 for generating a signal having a predetermined pitch period, and an unvoiced sound. An amplitude value generation unit 2 that indicates the magnitude of the amplitude of the random phase speech unit waveform from the switching unit 3 with a random value at the timing of the periodic signal output from the period generation unit 21 during speech synthesis.
2, an integrating unit 23 that integrates the random phase speech unit waveform from the switching unit 3 with the random value from the amplitude value generating unit 22 at the time of voice synthesis of unvoiced sound, and an integrating unit 23 at the time of voice synthesis of unvoiced sound. A switching unit 24 that switches to select the speech unit waveform 1 itself from the switching unit 3 during voice synthesis of a voiced sound, and a waveform generation unit that switches the waveform from the switching unit 24. Superimposing unit 2 that superimposes at a timing shifted by a periodic signal from 21
5 is provided. Note that the amplitude value generation unit 22 is configured by, for example, a random number generator.

In such a configuration, as shown in FIG.
The period generating section 21 outputs a signal P indicating the timing at which the overlapping section 25 shifts and overlaps the waveform, and the amplitude value generating section 22 outputs the periodic signal P from the period generating section 21.
In other words, the random value W of the amplitude is updated and given to the integrating unit 23 at each timing when the waveforms are shifted and overlapped by the overlapping unit 25. The integrating unit 23 changes the amplitude of the random-phase speech unit waveform at random by integrating the random value W of the amplitude and the random-phase speech unit waveform from the switching unit 3, and changes this through the switching unit 24. To the overlapping section 25. In the overlapping section 25,
Superposition is performed while shifting the speech unit waveforms having not only a phase but also a random amplitude at the timing of the superposition by the periodic signal P from the period generating unit 21 to synthesize an unvoiced sound waveform. The synthesized voice of the unvoiced sound obtained in this way is obtained by shifting the speech unit waveforms having not only a phase but also a random amplitude at a predetermined timing and superimposing them. At this time, since the amplitude of the speech unit waveform changes randomly each time the waveform is superimposed, the influence on the spectral characteristics is smaller than that of the conventional speech synthesizer.

Further, in the configuration example of FIG.
Is a random signal generator 31 for generating a random signal
And a pitch cycle generating section 32 for generating a signal with a predetermined pitch cycle, and switching to select a random signal from the random signal generating section 31 during voice synthesis of unvoiced sound,
At the time of voice synthesis of voiced sound, a switching unit 33 that switches so as to select the pitch period signal from the pitch period generation unit 32
And an amplitude value generating unit 34 that indicates the magnitude of the amplitude of the random phase speech unit waveform from the switching unit 3 at a random signal timing from the random signal generating unit 31 at the time of synthesis of the unvoiced sound by a random value; Switching unit 3 during voice synthesis
The integration unit 35 integrates the random phase speech unit waveform from the random number from the amplitude value generation unit 34 with the random value from the amplitude value generation unit 34, and switches to select the waveform integrated from the integration unit 35 when unvoiced sound is synthesized. When the voice is synthesized, the switching unit 3
A switching unit 36 for switching to select the speech unit waveform 1 itself from the first unit, and a superposition unit 37 for superposing the waveform from the switching unit 36 at a timing shifted by the random signal from the random signal generation unit 31. are doing. In addition,
The amplitude value generator 34 is constituted by, for example, a random number generator.

In such a configuration, at the time of voice synthesis of unvoiced sound, the random signal generator 31 outputs a random signal indicating a timing for shifting and overlapping the waveforms in the superimposing section 37, and the amplitude value generating section 34 The random value of the amplitude is updated by the random signal, that is, at each timing of superimposing the waveform while being shifted by the superimposing unit 37, and the updated random value is provided to the integrating unit 35. The integrating unit 35 changes the amplitude of the random-phase speech unit waveform at random by integrating the random value of the amplitude and the random-phase speech unit waveform from the switching unit 3, and changes the amplitude through the switching unit 36. It is sent to the overlapping section 37. The superposition unit 37 performs superposition while shifting the speech unit waveform having not only a phase but also a random amplitude at the timing of superposition by the random signal from the random signal generation unit 31,
Synthesize unvoiced speech waveform. In this way, the obtained synthesized voice of unvoiced sound is obtained by superimposing speech unit waveforms having not only a phase but also a random amplitude at random timing.

The random signal generator 11 as shown in FIG. 6 can be applied not only to the speech synthesizer of the present invention but also to a conventional speech synthesizer. For example, when the random signal generating unit 11 is applied to the waveform superimposing unit of the above-described conventional speech synthesizer in which the amplitude of the speech unit waveform and the superimposing cycle are randomly designated by random numbers, the random signal generating unit 11 has a larger size than the conventional one. In addition, when the random amplitude speech unit waveforms are superimposed, the overlap becomes more uniform, and the randomness of the phase characteristics can be further improved.

[0035]

As described above, according to the first aspect of the present invention, when voice synthesis of unvoiced sound is performed, the phase is randomized instead of the amplitude of the voice unit waveform, so that the synthesized voice obtained as the synthesized voice is obtained. Regarding the spectral characteristics of unvoiced sound,
While maintaining the spectral characteristics of the unvoiced sound actually uttered by humans, only the phase characteristics can be made sufficiently random like white noise, making it more natural and closer to the human voice than before. It is possible to generate an unvoiced synthesized voice.

According to the second aspect of the present invention, not only the phase of the speech segment waveform but also the timing of superposition are made random, so that the phase characteristics are further randomized and more similar to human voice. Natural synthesized speech can be generated.

According to the third aspect of the present invention, since windowing processing means for performing windowing processing on the speech unit waveform after random phase conversion is provided, discontinuity at the start and end is reduced. Can be superimposed while shifting the random phase speech unit waveform at a predetermined timing,
As a result, it is possible to generate a natural synthesized speech that is closer to the real human voice than the discontinuity is reduced.

[0038]

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of a speech synthesizer according to the present invention.

FIG. 2 is a diagram illustrating a configuration example of a random phase shift unit.

FIGS. 3A, 3B, and 3C are diagrams illustrating an example of superposition of waveforms in a waveform superposition unit. FIGS.

FIG. 4 is a diagram illustrating a configuration example of a waveform superimposing unit.

FIGS. 5A, 5B, and 5C are diagrams illustrating an example of a windowing process.

FIG. 6 is a diagram illustrating a configuration example of a random signal generation unit.

FIG. 7 is a diagram showing an example of the timing of superposition of random phase speech unit waveforms in the present invention.

FIG. 8 is a diagram illustrating a configuration example of a waveform superimposing unit.

FIG. 9 is a diagram for explaining a process of updating a random value of amplitude.

FIG. 10 is a diagram illustrating a configuration example of a waveform superimposing unit.

FIG. 11 is a diagram showing an example of the timing of superposition of speech unit waveforms in a conventional speech synthesis device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Speech unit waveform 2 Random phase conversion part 3 Switching part 4 Waveform superposition part 5 Fourier transformation part 6 Random number generation part 11 Random signal generation part 12 Pitch period generation part 13 Switching part 14 Superposition part 15 Window processing part 51 Random number generation Device 52 difference device 53 adder

Claims (3)

(57) [Claims] 1. A voice synthesizing apparatus for synthesizing voice segment waveforms by superimposing them, wherein a random phase generator generates a random phase voice unit waveform by randomizing a phase of the voice unit waveform, and the random phase generator. And a waveform superimposing means for adding or superimposing the random-phase speech unit waveforms generated by the above while shifting to synthesize an unvoiced speech waveform. 2. The speech synthesizer according to claim 1, wherein:
The waveform superimposing means includes a random signal generating means for instructing a superposition timing of a random-phase speech unit waveform by a random value, and a random-phase speech unit based on the superposition timing instructed by the random signal generating means. A speech synthesizing device comprising: superimposing means for synthesizing an unvoiced speech waveform by shifting and adding or superimposing the waveforms. 3. The speech synthesizer according to claim 1, wherein
The speech synthesis apparatus further includes windowing processing means for performing windowing processing on the speech unit waveform after randomization.