JPH0916198A - Excitation signal generating device and excitation signal generating method in low bit rate vocoder - Google Patents

Abstract

PURPOSE: To provide a low bit rate vocoder in which a range of a reproduced voice is widened and tone can be improved. CONSTITUTION: An excitation signal generating device in a low bit rate vocoder calculates low frequency band energy value E included in a voice input signal in a low frequency band energy calculating means 10. A mixing ratio in which a pulse train signal and a white noise signal are mixed based on calculated low frequency band energy value E is obtained in a mixing ratio calculating means 12. An amplification factor of an amplifier 16 and a coefficient (a) of a first FIR filter 14 are switched based on this mixing ratio, and the pulse train signal and the white noise signal are added in an adder 20 with a desired mixing ratio. After normalization is performed in an amplifier 22 to adjust an effective value of a synthesized voice to an actual voice input signal, a final mixing excitation signal is outputted. Since an excitation signal is generated by mixing the white noise signal and the pulse train signal, tone of a reproduced voice can be improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech synthesis by a low bit rate vocoder.

[0002]

2. Description of the Related Art There is known a method of synthesizing a voice by estimating a formant of a voice input signal. For example, FIG. 3 is a block diagram of a voice generation model described on page 141 of “Digital signal processing of voice (lower), translation: Kuki Suzuki, Corona Publishing”. According to the model shown in FIG. 3, the voiced sound is represented by the pitch period, the amplitude, and the three formant frequencies on the low frequency side, and the unvoiced sound is simply represented by the amplitude, zero, and the pole. In this way, a voice generation model is constructed based on the formant of the voice signal, and a pulse train signal having a predetermined pitch period is supplied by an impulse generator and a predetermined white noise is supplied by a white noise generator. As a result, the original voice signal is synthesized.

As described above, only the pulse train signal or the white noise signal is switched and supplied to the conventional synthesized sound of this type to excite the sound. And each single pulse train reproduces voiced sound and white noise reproduces unvoiced sound.

[0004]

As described above, in the conventional vocoder, a simple single pulse excitation signal (white noise in the case of unvoiced sound) is used for excitation when synthesizing a voice. However, it is difficult to synthesize all voices with such a simple single pulse train signal (or white noise). In addition, since it is extremely difficult to completely distinguish human voices from voiced sounds or unvoiced sounds, it is very difficult to distinguish between voiced sounds and unvoiced sounds. Had certain limits.

The present invention has been made to solve the above problems, and an object of the present invention is to improve the sound quality when synthesizing a voiced sound in speech synthesis by a low bit rate vocoder. .

[0006]

In order to solve the above-mentioned problems, a first aspect of the present invention provides a low bit rate vocoder which estimates a formant of a voice input signal and synthesizes an original voice input signal based on this formant. In the excitation signal generator used, based on the low frequency energy value obtained by the low frequency energy calculation means for obtaining the low frequency energy value which is the energy of the low frequency component of the voice input signal, and the low frequency energy calculation means. Then, the pulse train signal and the noise signal are mixed based on the mixture ratio calculating means for obtaining the mixture ratio of the pulse train signal and the noise signal, and the mixture ratio obtained by the mixture ratio determining means to generate the mixed excitation signal. And a mixed excitation signal generating means. For example, a white noise signal is suitable as the noise signal.

In order to solve the above-mentioned problems, a second aspect of the present invention is the excitation signal generator according to the first aspect of the present invention, wherein the mixing ratio calculating means includes the low-frequency energy value and a predetermined threshold value. Is compared, and the mixing ratio is calculated based on the result of this comparison.

In order to solve the above-mentioned problems, a third aspect of the present invention is the excitation signal generator according to the first aspect of the present invention, wherein the mixed excitation signal generating means includes a low frequency component of the pulse train signal, and It is an excitation signal generator characterized by mixing with a high frequency component of a noise signal.

In order to solve the above problem, a fourth aspect of the present invention is an excitation signal generation method used in a low bit rate vocoder for estimating a formant of a voice input signal and synthesizing an original voice input signal based on the formant. In the low-pass energy calculation step for obtaining a low-pass energy value that is the energy of the low-pass frequency component of the voice input signal,
Based on the low band energy value obtained in the low band energy calculation step, based on the mixing ratio obtained in the mixing ratio calculation step for obtaining the mixing ratio of the pulse train signal and the noise signal, the mixing ratio determination step, And a mixed excitation signal generating step of generating a mixed excitation signal by mixing the pulse train signal and the noise signal, and the excitation signal generating method.

In order to solve the above-mentioned problems, a fifth aspect of the present invention is the excitation signal generating method according to the first aspect of the present invention, wherein the mixing ratio calculation step comprises the low-range energy value and a predetermined threshold value. And a mixture ratio is calculated based on the result of this comparison.

A sixth aspect of the present invention, in order to solve the above-mentioned problems, in the excitation signal generating method according to the first aspect of the present invention, the mixed excitation signal generating step includes: a low-frequency component of the pulse train signal; A method of generating an excitation signal, which is characterized by mixing a high frequency component of a noise signal.

[0012]

The mixed excitation signal output means in the first aspect of the present invention mixes the pulse train signal and the noise signal based on the mixing ratio, and thereby generates the mixed excitation signal. Therefore, more detailed excitation is possible as compared with the case where only the pulse train signal or only the noise signal is used as the excitation signal.

The mixing ratio calculating means in the second invention is
The mixture ratio is calculated by comparing a predetermined threshold value and the low band energy value. Therefore, the mixing ratio can be easily calculated.

The mixed excitation signal generating means in the third aspect of the present invention mixes the low frequency component of the pulse train signal and the high frequency component of the noise signal. Therefore, a spectrally flat mixed excitation signal is obtained.

The fourth to sixth inventions are excitation signal generating methods corresponding to the first to third excitation signal generators, and their operation is the same as that of the first to third inventions. It is the same.

[0016]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an excitation signal generator in a low bit rate vocoder which is a preferred embodiment of the present invention. As shown in FIG. 1, the voice input signal is first converted into the low frequency energy calculation means 10
Supplied to The low frequency energy calculation means 10 calculates the energy of the low frequency components of the voice input signal. For example, the low frequency energy calculating means 10 calculates the low frequency energy value by rescuing only the signal component of the frequency band of 100 to 800 Hz of the audio input signal and calculating the energy of the signal component. The low frequency energy value is obtained by calculating the moving average of the low frequency component signals, but other average value calculation methods are also suitable.

The low band energy value calculated by the low band energy calculating means 10 in this manner is supplied to the mixture ratio calculating means 12. The mixing ratio calculating means 12 calculates the mixing ratio of the pulse train signal and the white noise signal based on the low band energy value.

A feature of this embodiment is that when a voiced sound is synthesized in a low bit rate vocoder, not only a pulse train signal is used, but a pulse train signal and a white noise signal are appropriately mixed to generate an excitation signal. That is what was generated. In this way, by mixing the pulse train signal and the white noise signal, it is possible to widen the range of the generated voice, improve the audibility, and improve the sound quality.

The pulse train signal used to generate the excitation signal is supplied to the first FIR filter 14. The white noise signal used for generating the excitation signal is supplied to the second FIR filter 18 via the amplifier 16. As shown in FIG. 1, the output signals of the first FIR filter 14 and the second filter 18 are added in an adder 20 and finally after a predetermined normalization in an amplifier 22, It is output to the outside. The configuration from the first FIR filter 14 to the amplifier 22 corresponds to the mixed excitation signal output means in the claims.

Next, the operation of generating the excitation signal in this embodiment will be described with reference to FIG. The excitation signal generated in this embodiment is called a mixed excitation signal.

FIG. 2 is a flowchart showing the operation of the excitation signal generator in the low bit rate vocoder according to this embodiment. As shown in the flowchart of FIG. 2, first, in step ST2-1, a 100 Hz high-pass filter is passed in order to remove the DC component of the input audio signal. As a result, the DC component is removed from the audio input signal.

Next, in step ST2-2, 800
Pass the low pass filter of Hz. by this,
Only the low frequency component of the audio input signal is extracted. The signal of 100 to 800 Hz thus obtained is called (x).

Next, the moving average of this signal (x) is calculated, and the average low frequency voiced energy LBVE is calculated. This calculation is performed by a first-order recursive filter. Here, the analysis frame of this recursive filter is the input signal of the frame energy in the voiced state. The above LBVE calculation is performed in step ST2-3.

Next, the obtained low band energy value is normalized (step ST2-4). The normalized low frequency energy value E is obtained by the following equation.

[0026]

## EQU1 ## E = (Σ | xi |) / (LBVE) (1) Here, E is the low frequency energy value normalized as described above, and xi is (x) described above. Is equivalent to.

Steps ST2-1 to ST2 described above
The operation up to -4 is performed in the low band energy value calculation means 10. The normalized low range energy value E is output from the low range energy calculation means 10, and this low range energy value E is supplied to the mixture ratio calculation means 12.

Next, in the mixing ratio calculating means 12, the low frequency energy value E is compared with a predetermined threshold value (ST2).
-5).

Then, as a result of this comparison, it is determined whether the low-frequency energy value is above the threshold value or below the threshold value.
Checked at T2-6. As a result of this comparison, when the low frequency energy value E is equal to or more than the threshold value, the process proceeds to step ST2-7, and the pulse train signal and the white noise signal are mixed at 80%: 20%. It is determined. On the other hand, when the low band energy value E is equal to or less than the threshold value, it is determined to mix the pulse train signal and the white noise signal with 50%: 50% (ST2-8).

As described above, in this embodiment, the mixing ratio is determined simply by the energy of the low frequency component of the voice input signal regardless of whether the voice frame is voiced or unvoiced. This is based on the empirical fact that voiced sound generally has a low-frequency energy of 30 dB or more higher than unvoiced sound. That is, it is known that the low-frequency energy value E in the case of unvoiced sound is a value sufficiently smaller than the low-frequency energy value E of voiced sound, and therefore it is suitable for distinguishing the voiced sound from the unvoiced sound. By setting such a threshold value, the mixing ratio is determined by simply calculating the low band energy value.

In this way, when the mixing ratio is determined in step ST2-7 or step ST2-8,
The determination about the mixing ratio in step ST2-9 is stored in “mix ratio flag”. The mixing ratio stored in this “mix ratio flag” is determined by the first FIR filter 14 as described later.
And the amplifier 16 respectively. In this embodiment, since the mixing ratio is either 80% or 50%,
As the “mix ratio flag”, a 1-bit configuration is sufficient. In the present embodiment, the mixing ratio was switched to two stages, but it is also possible to set it to three stages or more, and it goes without saying that the more stages there are, the finer the control can be performed.

Steps ST2-5 to ST
The operations up to 2-9 are performed by the mixture ratio calculation means 12. The “mix ratio flag” is also provided inside the mixing ratio calculating means 12. This "mix
The "ratio flag" data is supplied to the first FIR filter 14 and the amplifier 16 as shown in FIG.

The first FIR filter 14 is a first-order FIR filter for inputting a pulse train signal, and its transfer function is 1 + az ^-1 . This first FIR filter 1
4 is “mix r output from the mixture ratio calculation means 12
The coefficient a is switched based on the content of "audio flag". The amplifier 16 is an amplifier that inputs the white noise signal as described above, and its amplification factor is represented by G. The amplification factor G of the amplifier 16 is the mixture ratio calculation means 1
It is switched based on the value of the "mix ratio flag" output from 2.

In the present embodiment, the pulse train signal is extracted by the first FIR filter 14 as a low frequency signal only, while the white noise signal is extracted by the second FIR filter 18 as a high frequency component only. Be done. Then, the low-frequency signal of the pulse train signal and the high-frequency component of the white noise signal are added by the adder 20, normalized by the amplifier 22, and then output. The transfer function of the second FIR filter 18 is 1-bz
It is represented by ^-1 .

Next, the coefficient a of the first FIR filter 14
How the amplification factor G of the amplifier 16 and the coefficient b of the second FIR filter 18 are set will be described.

From the configuration block diagram shown in FIG.
It will be appreciated that the mixed excitation signal e (n) in this example is given by:

[0037]

## EQU00002 ## e (n) = G (Hw (n) * w (n)) + (Hp (n) * P (n)) (2) Here, the mixed excitation signal in the present embodiment is e Is represented by (n), P (n) represents a single periodic pulse train signal,
w (n) represents a white noise signal, respectively. Also, Hp
(N) represents the low frequency component filtered by the first FIR filter 14, and Hw (n) represents the high frequency component filtered by the second FIR filter 18. Further, G is the amplifier 1 as described above.
The amplification factor is 6.

When the pulse train signal and the white noise signal are white and independent and have a unified power spectrum, the power spectrum of the mixed excitation signal is calculated by the following formula. To be done.

[0039]

Se (w) = G ² | Hw (w) | ² + | Hp (w) | ² = G ² (Hb ² −2bcosw) + 1 + a ² + 2acosw (3) This equation (3) is converted into a spectrum. In order to obtain a substantially flat excitation signal, the constant of the frequency-dependent term may be set to 0, and each coefficient may satisfy the following equation.

[0040]

[Equation 4] 0 = G ² (−2bcosw) + 2acoswa a = bG ² (4) In the present embodiment, each coefficient is calculated as follows based on the relationship of the coefficients obtained by the equation (4). By switching, the above-mentioned mixing ratio of 80% and 50% is realized.

First of all, G could not recover two kinds of values less than or equal to 1, mixing ratio calculation *** This part could not be restored. ***
Means that. Further, in the present embodiment, the second FI
The coefficient b in the R filter 18 is fixed to 1.
As a result, the coefficient a of the first FIR filter 14 is calculated as A = G ² based on the above-mentioned equation (4). That is, in this embodiment, a = 0.0625 when the mixing ratio is 80%, and a = 1.0 when the mixing ratio is 50%.

A feature of this embodiment is that the pulse train signal and the white noise signal are mixed at a predetermined ratio based on the low-range energy value E. Since the excitation signal is generated by such mixing, it is possible to improve the sound quality of voiced sound. Further, the pulse train signal is passed through a low-pass filter (first FIR filter 14) to extract only low-frequency components, while a white noise signal is added with a high-pass filter (second FIR filter 18) to obtain high-frequency signals. Only the ingredients were taken out. By adding these low frequency components and high frequency components, a spectrally flat excitation signal can be generated in this embodiment.

In this way, the effective value of the excitation signal obtained by mixing the pulse train signal and the white noise signal is the voice input signal analyzed by the voice input signal analyzer not shown in FIG. Is compared to the effective value of. As a result of this comparison, normalization of the mixed excitation signal is performed in the amplifier 22. That is, this normalization is performed by setting the amplification factor γ of the amplifier 22 to a predetermined value. The mixed excitation signal finally obtained in this way is output to the synthesis filter. In the present embodiment, it was shown that the pulse train signal and the white noise signal are mixed in the case of voiced sound, but in the case of unvoiced sound, the excitation signal is generated using only the white noise signal as in the conventional case. It

[0044]

As described above, according to the first aspect of the present invention, since the voiced sound is synthesized by using the mixed excitation signal, it is possible to expand the range of the synthesized voice and improve the listening interval. A vocoder designed is obtained. As a result, a low bit rate vocoder with significantly improved sound quality can be realized.

In the second aspect of the present invention, the mixing ratio is calculated by comparing the low range energy value with a predetermined threshold value. As a result, the calculation of the mixing ratio becomes easy, and a low bit rate vocoder capable of improving sound quality with a simple configuration can be obtained.

According to the third aspect of the present invention, since the low frequency component of the pulse train signal and the high frequency component of the noise signal are mixed, a low bit rate vocoder capable of obtaining a spectrally flat mixed excitation signal is provided. can get.

The fourth to sixth excitation signal generation methods are as follows:
The method is realized by the excitation signal generator according to the first to third aspects of the present invention, and has the same effects as those of the first to third aspects of the present invention.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram showing an excitation signal generator for a low bit rate vocoder which is a preferred embodiment of the present invention.

FIG. 2 is a flowchart explaining the operation of the excitation signal generator shown in FIG.

FIG. 3 is an explanatory diagram of a speech synthesis model according to conventional formant estimation.

[Explanation of symbols]

10 low band energy calculating means, 12 mixing ratio calculating means, 14 first FIR filter, 16 amplifier, 18
Second FIR filter, 20 adder, 22 amplifier.

Claims (6)

[Claims] 1. An excitation signal generator used in a low bit rate vocoder for estimating a formant of an audio input signal and synthesizing an original audio input signal based on the formant, wherein an energy of a low frequency component of the audio input signal. A low range energy calculating means for obtaining a low range energy value that is, based on the low range energy value obtained by the low range energy calculating means, a mixture ratio calculating means for obtaining a mixture ratio of a pulse train signal and a noise signal, An excitation signal generation device comprising: a mixture excitation signal generation unit that mixes a pulse train signal and a noise signal based on the mixture ratio obtained by the mixture ratio determination unit to generate a mixed excitation signal. 2. The excitation signal generator according to claim 1, wherein the mixing ratio calculation means compares the low frequency energy value with a predetermined threshold value,
An excitation signal generator, wherein the mixing ratio is calculated based on the result of this comparison. 3. The excitation signal generator according to claim 1, wherein the mixed excitation signal generation means mixes a low frequency component of the pulse train signal and a high frequency component of the noise signal. Excitation signal generator. 4. An excitation signal generating method used in a low bit rate vocoder for estimating a formant of a voice input signal and synthesizing an original voice input signal based on the formant, wherein the energy of a low frequency component of the voice input signal. A low range energy calculating step for obtaining a low range energy value that is, based on the low range energy value obtained in the low range energy calculating step, a mixture ratio calculating step for obtaining a mixture ratio of a pulse train signal and a noise signal, A mixed excitation signal generation step of mixing a pulse train signal and a noise signal to generate a mixed excitation signal based on the mixing ratio obtained in the mixing ratio determination step, the excitation signal generating method. 5. The excitation signal generating method according to claim 4, wherein the mixing ratio calculating step compares the low band energy value with a predetermined threshold value,
An excitation signal generating method, characterized in that the mixing ratio is calculated based on the result of this comparison. 6. The excitation signal generating method according to claim 4, wherein the mixed excitation signal generating step mixes a low frequency component of the pulse train signal and a high frequency component of the noise signal. Excitation signal generation method.