JP3319556B2 - Formant enhancement method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a formant emphasis method capable of reducing an indistinct feeling with less processing amounts. SOLUTION: In the formant emphasis method emphasizing a formant being a peak part of a spectrum of an input sound signal and attenuating the trough part, the processing emphasizing the formant of the input sound signal and attenuating the trough part is performed by a spectrum emphasis filter 1101, and a spectrum slope incorporated in the output signal of the spectrum emphasis filter 1101 is corrected by a first characteristic variable filter 1103 adaptively varying its characteristic according to the characteristic of the input sound signal and a first characteristic fixed filter 1104.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a decoding section or a preprocessing section for speech processing in speech encoding / decoding processing, which emphasizes peaks of the spectrum of an input speech signal.
The present invention relates to a formant enhancement method for attenuating a valley portion.

[0002]

2. Description of the Related Art A technology for encoding voice signals at a low bit rate with high efficiency is an important technology for effective use of radio waves and reduction of communication costs in mobile communications such as car telephones and in-company communications. is there. As a speech encoding method capable of producing speech with excellent quality at a bit rate of 8 kbps or less,
A CELP (Code Excited Linear Prediction) system is known. This CELP method is used by AT & T Bell Labs MR
“Code-Excited Linea” by Schroeder and BSAtal
r Prediction (CELP) High-Quality Speech at Very Low
Bit Rates ”, Proc. ICASSP; 1985, pp. 937-939” (literature 1), which has attracted attention as a method for synthesizing high-quality speech. Considerations have been made. However, at a very low bit rate of 8 kbps or less, quality degradation of synthesized speech is perceived and is still insufficient.

Against this background, AT & T Bell Labs' P.I.T. technique improves the subjective quality by emphasizing peaks (formants) of the synthesized speech spectrum and performing post-processing to attenuate valleys. “Quantization Procedures for the Excitation in CEL” by Kroon and BSAtal et al.
P Coders ", Proc. ICASSP; 1987, pp. 1649-1652" (Reference 2). In Document 2, the quality is improved by using an all-pole filter configured by multiplying the LPC coefficient (linear prediction coefficient) sent to the decoder by a coefficient so as to smooth the spectral envelope in the post-processing. . Expressing this all-pole filter in the z-transform domain results in equation (1).

[0004]

(Equation 1)

Here, A (z / β) is expressed by equation (2).

[0006]

(Equation 2)

However, the all-pole type filter Q ₁ (z) has a problem that it includes an extra spectral gradient and often synthesizes a synthesized sound. JP-A-64-13
No. 200, entitled "Improvement in Method for Compressing Digitally Coded Speech" (Reference 3),
A formant enhancement filter that solves this problem is disclosed. This document 3 proposes a method of cascade-connecting a pole-zero type filter taking into account spectral tilt correction and a first-order high-pass filter having fixed characteristics. Expressing the transfer function of this formant enhancement filter in the z-transform domain results in equation (3).

[0008]

(Equation 3)

According to this formant emphasis filter,
The term A (z / γ) and the term (1−μz ⁻¹ ) in the equation (3) are term A.
Since it works so as to correct the extra spectral tilt of (z / β), the problem that the synthesized sound is muffled is solved. However, the filter order of the formant emphasis filter Q ₂ (z) is 2P + 1 order, and there is a problem that the processing amount increases.

Another formant enhancement filter is disclosed in Japanese Patent Application Laid-Open No. 2-82710, "Post-processing filter" (Reference 4).
Is disclosed. This document 4 proposes a pole-zero filter having a spectral tilt correction unit with a reduced order in the numerator term. If the transfer function of this formant enhancement filter is expressed in the z-transform domain, it is as shown in equation (4).

[0011]

(Equation 4)

The numerator A ^(M) (z / β) in the equation (4) serves to correct the spectrum tilt. In this case, although the processing amount decreases as the order M decreases, M must be increased to some extent in order for the spectrum tilt correction to function sufficiently. If M = 1, the formant emphasis filter will still cause muffledness.

A common problem of the equations (3) and (4) is that the filter coefficient of the formant emphasis filter is controlled by fixed values β and γ, or only β. For this reason, fine adjustment of the filter characteristic of the formant emphasis filter cannot be performed, and the sound quality improvement ability by the formant emphasis filter is limited. In addition, in formulas (3) and (4), the formant emphasis filter is always controlled using fixed β and γ, so that formant emphasis is strongly applied to a certain part of the input voice and weakened to a certain part. Adaptive processing could not be performed.

[0014]

As described above, in the conventional formant emphasis filter, the all-pole filter defined by the equation (1) has a problem that the synthesized sound is muffled and the subjective quality is reduced. . In addition, the cascade connection of the pole-zero type filter and the first-order high-pass filter defined by the equation (3) eliminates the muffled feeling of synthesized sound and improves the subjective quality, but increases the processing amount. was there. Further, in the pole-zero filter defined by the equation (4), when the processing amount is reduced by setting the order of the numerator M = 1,
Since spectral tilt correction cannot be performed sufficiently, there remains a problem that the muffled feeling of the synthesized sound cannot be eliminated.

Further, in the conventional formant emphasis filter, since the filter coefficient is controlled only by a fixed value β, γ or β, fine adjustment of the filter cannot be performed, and the sound quality improvement ability by the formant emphasis filter is limited. In addition, since the formant emphasis filter is always controlled using fixed β and γ, there is a problem that adaptive processing such as applying a strong formant emphasis to a certain part of the input voice and weakening it in a certain part cannot be performed.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and has as its object to provide a formant emphasis method capable of obtaining high-quality speech.

More specifically, it is an object of the present invention to provide a formant emphasizing method capable of obtaining high-quality sound with reduced muffled feeling with a small amount of processing.

It is another object of the present invention to provide a formant emphasis method that can control the filter coefficient of the formant emphasis filter finely to obtain higher-quality sound.

[0019]

According to a first aspect of the present invention, there is provided a formant emphasis method for emphasizing a peak portion of a spectrum of an input voice signal and attenuating a valley portion. And correcting the spectral gradient caused by this formant enhancement processing with a primary filter whose characteristics are adaptively changed according to the characteristics of the input audio signal or the spectral enhancement characteristics and a primary filter whose characteristics are fixed. Features.

A formant emphasis filter according to a first aspect of the present invention includes a main filter for performing a formant emphasis process for emphasizing a peak portion and attenuating a valley portion of a spectrum of an input audio signal, and an output signal of the main filter. A first and a second inclination correction filter for correcting the spectral inclination to be applied are cascaded, and the first spectral inclination correction filter is adaptively adapted according to the characteristics of the input audio signal or the characteristics of the spectrum emphasis filter. , And the second spectral tilt correction filter is a first-order filter having fixed characteristics.

In the first invention, the characteristic of the first spectral tilt correction filter is determined based on an input speech signal or a primary PARCOR coefficient (partial autocorrelation coefficient or partial autocorrelation coefficient) obtained from the main filter. May be provided.

Further, in the first invention, there is provided means for restricting and determining the characteristic of the first spectral tilt correction filter so as not to exceed a range determined from the characteristic obtained one unit time ago. Is also good.

As described above, in the first aspect, the characteristic of the input audio signal is corrected in order to enhance the peaks of the spectrum of the input audio signal and to correct the extra spectrum inclination generated by the main filter that attenuates the valleys. Alternatively, rough spectral tilt correction is performed by a first spectral tilt correction filter including a first-order filter whose characteristics are adaptively changed according to the characteristics of the main filter. Since the first spectral tilt correction filter has the first order, it can be realized by slightly increasing the processing amount. Further, by passing the audio signal through a second spectral tilt correction comprising a primary filter having a fixed characteristic, an extra spectral tilt that cannot be completely removed by the first spectral tilt correction filter is corrected. Since the order of the second spectral tilt correction filter is also the first order, the correction can be performed with little increase in the processing amount.

That is, for example, the formant emphasis filter defined by the equation (3) requires 2P + 1 times of product sum, whereas the present invention enables formant emphasis processing by P + 2 times of product sum, and reduces the processing amount to almost half. Can be.

Further, the extra spectral gradient included in the main filter for emphasizing the peaks and attenuating the valleys of the spectrum of the input voice signal is a simple spectral characteristic that can be almost expressed by a first-order filter. is there. For this reason,
As in the present invention, the use of the first-order characteristic variable filter and the first-order characteristic fixed filter can sufficiently and effectively correct an extra spectral tilt. For example, in the case of the spectrum tilt correction of the conventional example represented by the equation (3), the correction can be performed with high accuracy due to the large order, but the spectral characteristics of the extra spectrum tilt included in the main filter are simple. Therefore, the correction can be sufficiently performed even in the cascade connection of the primary characteristic variable filter and the primary characteristic fixed filter, and there is almost no audible difference compared with the conventional method. further,
In the formant emphasis filter defined by the expression (4), if the order of the numerator is set to M = 1, the number of times of product sum is almost the same as that of the present invention, but the effect of the spectral tilt correction cannot be sufficiently obtained. On the other hand, in the present invention, since the primary filter whose characteristic is variable and the primary filter whose characteristic is fixed are connected in cascade, the spectrum tilt can be corrected sufficiently effectively.

As described above, according to the first aspect of the present invention, the formant emphasis filter is constituted by the main filter, the primary characteristic spectrum inclination correction filter having variable characteristics, and the primary characteristic spectrum inclination correction filter having fixed characteristics. Formant emphasis processing that does not cause a muffled feeling can be performed by the amount, and the subjective quality is effectively improved.

[0027] The formant enhancement method according to the second invention comprises:
Formant enhancement processing for enhancing peaks of the spectrum of the input audio signal and attenuating valleys is performed by a polar filter, and processing for correcting the spectrum inclination caused by the formant enhancement processing is performed by a zero-type filter. At least one of the filter coefficient of the type filter and the filter coefficient of the zero-type filter is set to L of the input audio signal.
It is characterized in that it is determined by a product of a predetermined constant arbitrarily corresponding to each coefficient of the PC coefficient and each coefficient of the LPC coefficient.

A formant emphasis filter according to a second aspect of the present invention includes a polar filter for performing a formant emphasis process for emphasizing a peak portion of a spectrum of an input audio signal and attenuating a valley portion, and an output signal of the pole filter. Filter means comprising a cascade connection of zero-type filters for correcting the spectral tilt contained in the filter, and filter coefficient determination means for determining the filter coefficients of the pole-type filter and the zero-type filter. The means has constant storage means for storing a plurality of constants arbitrarily determined in advance corresponding to respective coefficients of the LPC coefficient, and at least one of a filter coefficient of the polar filter and a filter coefficient of the zero-type filter. Is determined by the product of each next coefficient of the LPC coefficient of the input audio signal and the corresponding constant stored in the constant storage means. That. The constant storage means is constituted by a memory table.

As described above, in the second aspect, the filter coefficient is determined by the product of the LPC coefficient of the input audio signal and a plurality of arbitrarily predetermined constants corresponding to each of the following coefficients of the LPC coefficient. Therefore, the characteristics of the formant emphasis filter can be freely determined by setting the plurality of constants.

The conventional formant emphasis filter is expressed by (3)
As shown in the equation, the filter is composed of a polar filter having a transfer function of 1 / A (z / β) and a zero-type filter having a transfer function of A (z / γ). Thus, the strength of the formant emphasis can be determined. However, as can be seen from equation (2), the polar filter filter coefficients are expressed as {α _i β ⁱ ; i = 1 to P}, and similarly, the filter coefficients of the zero-type filter are also {α
_i γ ⁱ ; i = 1 to P｝, so that the LPC coefficients {α _i ; i
= 1 to P}, there is naturally a limitation that the exponential function value of β and γ can only be taken as {β ⁱ ; i = 1 to P}, {γ ⁱ ; i = 1 to P}. .

The purpose of the formant enhancement filter is to improve the subjective quality. However, whether or not the quality of the sound is subjectively improved is determined by repeating the trial listening of the reproduced sound signal and the adjustment of the parameters. Is common. For this reason, it is more advantageous to improve the quality of the sound by the formant emphasis filter if the coefficient multiplied by the LPC coefficient to obtain the filter coefficient is not limited to the exponential function value as in the prior art, but can be freely set as in the present invention. Becomes

In a formant emphasis method according to another aspect of the second invention, a plurality of types of constant storage means for storing a plurality of constants arbitrarily determined in advance corresponding to the respective following coefficients of the LPC coefficient are provided. At least one of the filter coefficient of the polar filter and the filter coefficient of the zero-type filter is selected from a plurality of types of constant storage means based on the following coefficients of the LPC coefficient of the input audio signal and the attribute of the input audio signal. It is determined by the product of the corresponding constants stored in the two constant storage means.

Originally, formants such as vowels appear strongly in voice signals, and regions where quality can be improved by emphasizing them, and clear formants such as consonants do not easily appear. There are areas where better results are obtained when the level is weaker, and the final subjective quality can be improved by adaptively changing the degree of emphasis according to the attribute of the input audio signal.

It is also effective to weaken the formant emphasis in, for example, a background portion where no voice exists, and to strengthen the formant emphasis in a region where voice exists. Signals in the background are often noises represented by, for example, engine sounds and air-conditioning sounds, and if such signals are subjected to spectrum emphasis and attenuation by conventional methods, they will be perceived as unnatural sounds. That is the main reason. The conventional method does not take any measure against such a problem.

On the other hand, in the second invention, a plurality of arbitrarily predetermined LPC coefficients are respectively associated with the following LPC coefficients so that the degree of the formant enhancement can be varied stepwise. A memory table, which is a plurality of types of constant storage means storing constants, is prepared in advance,
A memory table to be used is adaptively selected according to attributes such as a vowel part, a consonant part, and a background part of the input voice signal. Thereby, the memory table most suitable for the attribute of the input audio signal is always selected, and as a result, the quality of the audio after the formant enhancement is improved.

[0036]

Embodiments of the present invention will be described below with reference to the drawings.

[First Embodiment of the Invention] (First Embodiment) FIG. 1 is a block diagram for explaining a basic operation of a formant emphasis filter according to a first embodiment. In the figure, a digitally processed audio signal is sequentially input from an input terminal 1001 in frame units with a plurality of samples as one frame. In the present embodiment, a description will be given assuming that one sample is composed of 40 samples. On the other hand, LPC coefficients representing the spectral outline of each in-frame audio signal are sequentially input from an input terminal 1002. In the formant enhancement filter 1003, the input terminal 1
001 is subjected to formant emphasis using the LPC coefficient input from the input terminal 1002,
The resulting output signal is output from output terminal 1004.

FIG. 2 shows the formant enhancement filter 1 of FIG.
It is a block diagram which shows the internal structure of 003. The formant emphasis filter 1003 shown in FIG. 2 is configured by cascade-connecting a spectrum emphasis filter 1101, a characteristic variable filter 1103 whose characteristics are controlled by a filter coefficient determination unit 1102, and a characteristic fixed filter 1104.

The spectrum emphasis filter 1101 has a peak portion (formant) of the spectrum of the input audio signal.
, And functions as a main filter that achieves the basic function of the formant enhancement filter 1003 to attenuate the valley portion, and is configured based on the LPC coefficient obtained from the input terminal 1002. The configuration of the spectrum emphasis filter 1101 is determined by using LPC coefficients {α _i ; i = 1 to P}.
When expressed in the z-transform domain using, the following equation (5) is obtained.

[0040]

(Equation 5)

Where C (z) is the z-converted value of the input audio signal, and E (z) is the z-converted value of the output signal. Further, P represents a filter order, and in the present embodiment, P = 10 will be described. Here, β is a constant (0 <β <1) for controlling the degree of spectrum intensity. As β approaches 1, the degree of spectrum emphasis increases and the effect of noise suppression increases, but there is an adverse effect that the muffled feeling of the synthesized sound also increases. Conversely, the degree of spectrum enhancement decreases as β approaches 0, and the effect of noise suppression also decreases.

When equation (5) is expressed in the time domain, equation (6) is obtained.

[0043]

(Equation 6)

Here, c (n) is expressed as a time domain signal of C (z), and e (n) is expressed as a time domain signal of E (z).

[0045] Based on the LPC coefficients given from the input terminal 1002, the filter coefficient mu ₁ is determined by the filter coefficient determining unit 1102. μ ₁ is determined so as to correct the spectral tilt inherent in the all-pole filter defined by the LPC coefficient. If all-pole filter defined by the LPC coefficients has a low-pass characteristic, mu ₁ takes a negative value,
If all-pole filter defined by the LPC coefficients conversely has a high-pass characteristic, mu ₁ takes a positive value. For mu ₁ specific determination method will be described later.

The output signal e (n) of the spectrum emphasis filter and the output μ ₁ of the filter coefficient determination unit 1102 are input to the characteristic variable filter 1103. Variable characteristic filter 1
The filter order of the filter 103 is the first order, and the output signal of the characteristic variable filter 1103 is expressed as F (z) in the z-transform domain.
F (z) is as shown in equation (7).

F (z) = (1 + μ ₁ z ⁻¹ ) E (z) (7) When the equation (7) is expressed in the time domain, the equation (8) is obtained. Here, e (n) is expressed as a time domain signal of E (z), and f (n) is expressed as a time domain signal of F (z).

F (n) = e (n) + μ ₁ e (n−1) (8) As can be seen from the equation (8), when the all-pole filter defined by the LPC coefficient has the high-pass characteristic, μ ₁ is a low-pass filter for taking a plus value, and works to correct the high-pass characteristic of the all-pole filter defined by the LPC coefficient. Conversely, mu ₁ when having the all-pole filter is a low-pass characteristic defined by the LPC coefficients serve as the high-pass filter for taking a negative value acts to correct the low-pass characteristic of the all-pole filter defined by the LPC coefficients .

Output f (n) of variable characteristic filter 1103
Is given to the characteristic fixed filter 1104. The filter order of the characteristic fixed filter 1104 is the first order, and the output signal of the characteristic variable filter 1103 is G (z) in the z-transform domain.
, G (z) is as shown in equation (9).

G (z) = (1−μ ₂ z ⁻¹ ) F (z) (9) When Expression (9) is expressed in the time domain, Expression (10) is obtained. Here, f (n) is expressed as a time domain signal of F (z), and g (n) is expressed as a time domain signal of G (z).

G (n) = f (n) −μ ₂ f (n−1) (10) Since μ ₂ is a positive fixed value, the characteristic fixed filter 1104 always has a high-pass characteristic according to Expression (9). . This is because the filter characteristic of the spectrum emphasis filter 1101 becomes a low-pass characteristic in most cases for voiced sounds that are perceptually important, and becomes a high-pass filter by the characteristic variable filter 1103 to correct this, but in many cases, it cannot be corrected completely. A feeling of muffledness remained. In order to solve this, a fixed characteristic filter 1104 having a high-pass characteristic is prepared. The output signal g (n) thus obtained
Is output from the output terminal 1004.

The flow of the above processing is summarized as a flowchart shown in FIG. Where ｛c (n),
n = 0 to NUM-1} is a digitally processed input audio signal, which represents a signal sequentially input from the input terminal 1001, and {e (n), n = -P to NUM-1} and {}. f (n), n = −1 to NUM−1} represents the internal state of the filter. {G (n), n = 0 to NUM-1} is an output audio signal, which is sequentially output from the output terminal 1004. When the variable n of e (n) and f (n) is a negative value, it means that the internal state of the previous frame is used. Here, NUM represents a frame length, and NUM = 4.
Set to 0. Further, P represents the order of the spectrum emphasis filter, and here, P = 10.

First, at step S11, the variable n is cleared to 0, and then at step S12, a spectrum emphasis process is performed to obtain e (n). Next, in step S13, a characteristic variable inclination correction process is performed to obtain f (n), and then in step S14, a characteristic fixed inclination correction process is performed to obtain g (n), and g (n) is output from the output terminal 1004. Output. In the next step S15, the variable n is increased by one, and in step S1
In step 6, the size of n is compared with the size of NUM. If the variable n is smaller than NUM, the process returns to step S12. If the variable n is equal to or larger than NUM, the process proceeds to step S17. Step S
At 17, the internal state of the filter is updated for the next frame, and the process ends in preparation for the input audio signal of the next frame.

In the above processing, step S12, step S13 and step S14 may be in any order. However, when the order is changed, it is natural that it is necessary to allocate an internal state that matches the order.

(Second Embodiment) FIG. 4 is a block diagram showing the configuration of the second embodiment. In FIG. 4, FIG.
Components having the same reference numerals as those having the same functions as those in FIG. Second Embodiment and First
The difference from this embodiment is that the input to the filter coefficient determination unit 1102 is different.

That is, the input to the filter coefficient determination unit 1102 in the present embodiment is a weighted LPC coefficient ｛α _i β ⁱ used in the spectrum enhancement filter 1101;
i = 1 to P｝. Since the weighted LPC coefficients are the filter coefficients of the spectrum emphasis filter 1101, the characteristics of the filter actually used for spectrum emphasis can be accurately obtained. In the present embodiment, since the seek filter coefficient mu ₁ of characteristic variable filter 1103 on the basis of the weighted LPC coefficients, thereby enabling more accurate spectral tilt correction.

FIG. 5 is a block diagram showing an example of the configuration of the filter coefficient determination section 1102. Either LPC coefficient {α _i ; i = 1 to P} or weighted LPC coefficient {α _i β ⁱ ; i = 1 to P} is input from input terminal 1404. Next, in a coefficient conversion unit 1401 that converts LPC coefficients into PARCOR coefficients (partial autocorrelation coefficients or partial autocorrelation coefficients), the input LPC coefficients or weighted LPC coefficients are converted into PARCOR coefficients. The specific method is described in "Digital Speech Processing" by Mr. Furui, Tokai University Press (Reference 5), and the description is omitted here. Coefficient conversion unit 1
401 PARCOR coefficient k ₁ of the primary is output from.

The following are known as properties unique to the PARCOR coefficient. In other words, if the spectrum of the filter composed of the LPC coefficients input to coefficient conversion section 1401 has a low-pass characteristic, the first-order PARCOR coefficient takes a negative value, and the stronger the low-pass characteristic, the higher the first-order PARCOR coefficient.
The ARCOR coefficient approaches −1. Conversely, if the spectrum has a high-pass characteristic, the first-order PARCOR coefficient takes a positive value, and the stronger the high-pass characteristic, the higher the first-order PARCOR coefficient.
The RCOR coefficient approaches +1. Therefore, the first PARC
By controlling the filter characteristic of the characteristic variable filter 1103 defined by the equation (7) using the OR coefficient, the coefficient conversion unit 1
It is possible to efficiently correct an LPC coefficient input to 401, that is, an extra spectrum tilt included in the spectrum outline of the spectrum enhancement filter 1101. Specifically, the multiplier 1402 multiplies the primary PARCOR coefficient k ₁ obtained from the coefficient conversion unit 1401 by a positive constant ε, and outputs the result from the output terminal 1403 as μ ₁ .

Μ ₁ = k ₁ ε (11) The above processing flow is summarized as a flowchart shown in FIG. Here, ｛c (n), n = 0 to NU
M-1} is a digitally processed audio signal that is sequentially input from the input terminal 1001, and {e}
(N), n = −P〜NUM−1} and {f (n), n
= −1 to NUM−1} represents the internal state of the filter.
{G (n), n = 0 to NUM-1} is an output signal,
The signals are sequentially output from the output terminal 1004. Also, e
When the variable n of (n) and f (n) is a negative value, it means that the internal state of the previous frame is used. Here, NUM represents the frame length, and NUM = 40. Further, P represents the order of the spectrum emphasis filter, and here, P = 10. Step S2 in FIG.
1, S22, S24, S25, S26 and S27
Steps S11, S12, S in FIG.
Since it is the same processing as 14, S15, S16 and S17,
Description is omitted.

The process newly provided in FIG. 6 is step S23. Characterized in the step S23 is that controlling the characteristic variable inclination correction characteristic in the first-order PARCOR coefficient k _1. That is, the product of the first-order PARCOR coefficient k ₁ and the constant ε is used as the filter coefficient of the first-order zero-type filter,
f (n).

In the above processing, step S22, step S23 and step S24 may be in any order. However, when the order is changed, it is natural that it is necessary to allocate an internal state that matches the order.

FIG. 7 is a block diagram showing another example of the configuration of the filter coefficient determination unit 1102. In FIG. 7, FIG.
The components denoted by the same reference numerals as those having the same functions as those in FIG. The filter coefficient determination unit 1102 shown in FIG. 5 and the filter coefficient determination unit 1 shown in FIG.
The difference of the configuration of the 102 is that is a limit as to fall within the scope of the filter coefficient mu ₁ output in the current frame is defined by mu ₁ value of the previous frame.

For this reason, in the configuration of FIG.
A buffer 1502 for storing ₁ is provided. When μ ₁ in the previous frame is expressed as μ ₁ p, μ ₁ p is used by the filter coefficient limiting unit 1501 to limit the fluctuation of μ ₁ . Μ ₁ of the current frame obtained as a result of the multiplication by the multiplier 1402 is input to the filter coefficient limiting unit 1501, and μ ₁ p stored in the buffer 1502 is also supplied to the filter coefficient limiting unit 1501 at the same time. In the filter coefficient limiting unit 1501 gives the limit so that the possible range of mu ₁ according to the following equation becomes _{μ 1 p-T ≦ μ 1} ≦ μ 1 p + T. Here, T is a positive constant.

Μ ₁ = μ ₁ p−T (if μ ₁ <μ ₁ p−T) (12) μ ₁ = μ ₁ p + T (if μ ₁ > μ ₁ p + T) (13) Equations (12) and (13) ) after giving the limit on mu ₁ according equation, and outputs a mu ₁ from the output terminal 1403. At the same time, μ ₁ is stored in buffer 1502 for the next frame.
It is stored as ₁ p.

By limiting the variation of the filter coefficient μ _{1 in} this manner, the characteristic of the characteristic variable filter 1103 does not greatly change, and the variation of the filter gain of the characteristic variable filter also decreases. Therefore, there is an advantage that the discontinuity of the gain between the frames is reduced, and abnormal noise is hardly generated.

The flow of the above processing is summarized as a flowchart shown in FIG. Where ｛c (n),
n = 0 to NUM-1} is a digitally processed audio signal, which represents a signal sequentially input from the input terminal 1001, and {e (n), n = -P to NUM-1} and {f (N), n = −1 to NUM−1} represents the internal state of the filter. {G (n), n = 0 to NUM-1} is an output signal, which is sequentially output from the output terminal 1004.
When the variable n of e (n) and f (n) is a negative value, it means that the internal state of the previous frame is used. Here, NUM represents the frame length, and NUM = 40. Further, P represents the order of the spectrum emphasis filter, and here, P = 10. Steps S37, S38, S39, S40, S41, S42 in FIG.
And S43 correspond to S11, S11 in FIG.
12, S13, S14, S15, S16, and S17 are the same processes, and thus description thereof is omitted.

The processes newly provided in FIG. 8 are steps S31 to S36. The feature of this processing is that the characteristic of the characteristic variable inclination correction processing is represented by a primary PARCOR coefficient k.
_{This is in} that the control is performed in step _{1 and} that the variation in the characteristics of the characteristic variable inclination correction processing is limited. Hereinafter, step S3
Steps 1 to 36 will be described.

In step S31, the primary PARCOR
The product of the coefficient k ₁ and the constant ε is obtained, and this value is stored in the variable μ ₁ . In step S32, performs a comparison between the variable mu ₁ and μ ₁ p-T, the smaller the better of mu ₁ step S3
The process proceeds to step S34 otherwise. In step S33, the value of mu ₁ is replaced with μ ₁ p-T proceeds to step S36. In step S34, the variables μ ₁ and μ ₁
performs a comparison between p + T, in step S35 is larger is more of μ _1, the process proceeds to step S36 otherwise.
In step S35, the value of mu ₁ is replaced with μ ₁ p + T proceeds to step S36. At step S36, it stores the value of mu ₁ for the next frame to mu ₁ p Step S37
Proceed to.

In the above processing, step S38, step S39 and step S40 may be in any order. However, when the order is changed, it is natural that it is necessary to allocate an internal state that matches the order.

(Third Embodiment) FIG. 9 is a block diagram showing a configuration of a formant enhancement filter according to a third embodiment. The difference between the third embodiment and the first embodiment is that the gain adjustment unit 1601 is included in the components.

The gain adjusting section 1601 adjusts the power of the output signal of the formant emphasis filter 1003 so as to match the power of the digitally processed audio signal which is the input signal of the formant emphasis filter 1003. A process for smoothly connecting the frames so that no discontinuity occurs between the frames and the current frame is performed. With this processing, even if the filter gain of the formant emphasis filter 1003 fluctuates greatly, the gain adjustment unit 1601
, The gain is adjusted, and generation of abnormal noise can be prevented.

(Fourth Embodiment) FIG. 10 is a block diagram showing a configuration of a formant emphasis filter according to a fourth embodiment, which is used together with a pitch emphasis filter 2202. 10, the components denoted by the same reference numerals as those in FIG. 9 have the same functions as those in FIG.

The pitch period L and the filter gain δ are input from the input terminal 2201 to the pitch emphasis filter 2202, and at the same time, the output signal g (n) of the formant emphasis filter 1003 is input. Pitch emphasis filter 2202
When the z-converted value of the input audio signal g (n) is G (z), the z-converted value V (z) of the output signal v (n) is expressed by the following equation.

[0074]

(Equation 7)

When this equation is expressed in the time domain, it becomes as equation (17).

V (n) = g (n) + δv (n−L) (17) The pitch emphasis filter 2202 performs pitch emphasis based on the expression (17), and outputs the output signal v (n) to the gain adjustment unit 160.
Give to 1.

As described above, when the pitch emphasis processing is performed in addition to the formant emphasis, noise suppression is more effectively performed, and there is a merit that the voice quality is improved. Here, the primary all-pole pitch emphasis filter has been described as the pitch emphasis filter 2202, but the present invention is not limited to this. The order of the formant emphasis filter 1003 and the pitch emphasis filter 2202 may be any order.

The recommended values of the respective constants of the present invention described above are as follows.

Β = 0.85, ε = 0.8, μ ₂ = 0.4, T = 0.3 These values are obtained experimentally by repeating trial listening, but depending on the taste of sound quality, Other settings are possible.
Of course, the present invention is not limited to this set value.

(Fifth Embodiment) As a fifth embodiment, FIG. 11 shows an embodiment in which the present invention is applied to a speech decoding device of a speech encoding / decoding system. 11, components denoted by the same reference numerals as those in FIG. 2 have the same functions as those in FIG.

In FIG. 11, a bit stream sent from a speech coding apparatus (not shown) via a transmission path is input to an input terminal 1701, and a demultiplexer 1
LSP coefficient index I by the bit operation in 702
LSP, adaptive codebook index IACB, noise codebook index ISCB, adaptive gain index IGA, and noise gain index IGS are output separately.

In the LSP coefficient inverse quantization unit 1703, LS
The LSP coefficient is decoded based on the P coefficient index ILSP. Next, the coefficient conversion unit 1712 calculates the decoded LSP coefficient as L
Convert to PC coefficient. The conversion method is described in the above-mentioned document 5, and the description is omitted here. The decoded LPC coefficient obtained in this way is
09 and formant enhancement filter 1003.

Next, an adaptive vector is selected from adaptive codebook 1704 using adaptive codebook index IACB as an index. Similarly, a noise vector is selected from the noise codebook 1705 based on the noise codebook index ISCB.

The adaptive gain decoding unit 1710 decodes the adaptive gain using the adaptive gain index IGA as an index, and the noise gain decoding unit 1711 similarly decodes the noise gain using the noise gain index IGS as an index.

The multiplier 1706 multiplies the adaptive vector by the adaptive gain, the multiplier 1707 multiplies the noise vector by the noise gain, the adder 1708 adds the output of the multiplier 1706 and the output of the 1707, and converts the driving vector. Generate. This drive vector is provided to synthesis filter 1709 and stored in adaptive codebook 1704 for processing of the next frame.

Assuming that the adaptive vector is f (n), the adaptive gain is a, the noise vector is u (n), and the noise gain is b, the driving vector c (n) is expressed by the following equation.

C (n) = a · f (n) + bu · n (n) (14) The synthesis filter 1709 performs filtering based on the decoded LPC coefficients obtained from the coefficient conversion unit 1712. When the decoded LPC coefficient is expressed as {α _i ; i = 1 to P} (P: filter order), the synthesis filter 1709 performs processing according to the following equation. Here, c (n) represents a drive vector as an input, and e (n) represents a composite vector as an output.

[0088]

(Equation 8)

Next, the synthesized vector e (n) and the decoded LPC coefficient ｛α _i ; i = 1
｝ P｝ is input and formant enhancement is performed as described above. The gain of the formant-emphasized signal is calculated using the gain of the composite vector e (n).
1 and the gain-adjusted signal is output to the output terminal 11.
Output from 04.

In FIG. 11, the configuration of FIG.
02 shows the case where the configuration of FIG. 6 is used,
Using the configuration of FIG. 4 for the formant enhancement filter 1003,
The configuration of FIG. 5 may be used for the filter coefficient determination unit 1102, and the combination of the formant enhancement filter 1003 and the filter coefficient determination unit 1102 included therein is arbitrary.

(Sixth Embodiment) As a sixth embodiment, FIG. 12 shows another embodiment in which the present invention is applied to a speech decoding device of a speech encoding / decoding system. 12, the components denoted by the same reference numerals as those in FIG.
The description is omitted assuming that it has the same function as 1.

The difference between the fifth embodiment and the sixth embodiment is that in the fifth embodiment, the LSP coefficient inverse quantization unit 1703 is used.
In the sixth embodiment, PARCO is used.
The point is that an R coefficient inverse quantization unit 1801 is used. Which coefficient is to be decoded depends on what coefficient has been encoded by a speech encoding device (not shown). That is, if the LSP coefficient is encoded by the audio encoding device, the LSP coefficient decoding unit 17036 is used in the audio decoding device. Similarly, in a speech encoding device, PARC
If the OR coefficient is encoded, the speech decoding device uses P
The ARCOR coefficient decoding unit 1801 will be used.

Next, the decoded PARCOR coefficient is converted using a coefficient conversion unit 1802 for converting the PARCOR coefficient to the LPC coefficient.
Convert from coefficients to LPC coefficients. This coefficient conversion unit 180
The concrete implementation method of the second is described in Reference 5, and the description is omitted here. The decoded LPC coefficients thus obtained are provided to the synthesis filter 1709 and the formant enhancement filter 1003. In the present embodiment, PA
Decoding PA as output of RCOR coefficient inverse quantization section 1801
Since the RCOR coefficient is obtained, it is not necessary to obtain the PARCOR coefficient again using the coefficient conversion unit in the filter coefficient determination unit 1102 as in the previous embodiments.
Decoding PA output from RCOR coefficient inverse quantization section 1801
The RCOR coefficient may be provided to the filter coefficient determination unit 1102. This simplifies the configuration and reduces the amount of processing.

In this embodiment, as shown in FIG. 13, the formant emphasis filter 1003 has an input terminal 100
1, an LPC coefficient is input from the input terminal 1002, and a PARCOR coefficient is input from the input terminal 1901. A formant-enhanced audio signal is output from the output terminal 1004. As in the present embodiment, the LP in the pre-processing unit of the formant enhancement filter 1003
In the case where the C coefficient and the PARCOR coefficient can be obtained, if both coefficients are given to the formant enhancement filter 1003, the coefficient conversion section 1401 in the filter coefficient determination section 1102 in the formant enhancement filter 1003 can be omitted from the configuration. it can.

In FIG. 12, the configuration of FIG. 2 is used for the formant enhancement filter 1003, and the filter coefficient determination unit 11
02 shows the case where the configuration of FIG. 7 is used,
Using the configuration of FIG. 4 for the formant enhancement filter 1003,
The configuration of FIG. 5 may be used for filter coefficient determination section 1102, and the combination of formant enhancement filter 1003 and filter coefficient determination section 1102 included therein is arbitrary.

(Seventh Embodiment) As a seventh embodiment, another embodiment in which the present invention is applied to a speech decoding apparatus of a speech encoding / decoding system is shown in FIG. In FIG. 14, components denoted by the same reference numerals as those in FIG.
The description is omitted assuming that it has the same function as 1.

The difference between the fifth and sixth embodiments and the seventh embodiment is that, in the fifth and sixth embodiments, the decoded LPC coefficients obtained by decoding in the formant emphasis filter 1003 by the decoding unit are sometimes different. Receives the decoded PARCO coefficient, whereas in the present embodiment, the synthesis filter 17
LPC analysis is performed on the output signal 09 to obtain a new LPC coefficient and a PARCOR coefficient in some cases, and formant enhancement is performed using the information. In the present embodiment, since the LPC coefficient of the composite signal is obtained again, the formant emphasis can be accurately performed. In addition, since the LPC analysis order can be set arbitrarily, when the analysis order is set to be large (analysis order> 10), the formant enhancement can be more finely controlled.

The LPC coefficient analyzer 2001 can perform analysis using the autocorrelation method or the covariance method.
In the autocorrelation method, the solution can be efficiently solved by using Durbin's recursive solution.
The coefficient and the PARCOR coefficient can be obtained simultaneously. Therefore, the formant enhancement filter 1003 includes L
What is necessary is just to give both a PC coefficient and a PARCOR coefficient. L
When the covariance method is used in the PC coefficient analysis unit 2001, Ch
If olesky decomposition is used, it can be solved efficiently. In this case, since only the LPC coefficient is obtained, only the LPC coefficient needs to be given to the formant enhancement filter 1003. FIG. 14 shows a configuration assuming a case where an LPC coefficient analysis unit 2001 based on the autocorrelation method is used.
This can also be realized using an LPC coefficient analysis unit based on the covariance method.

In FIG. 14, the configuration of FIG. 2 is used for the formant enhancement filter 1003, and the filter coefficient determination unit 11
02 shows the case where the configuration of FIG. 6 is used,
Using the configuration of FIG. 4 for the formant enhancement filter 1003,
The configuration of FIG. 5 may be used for the filter coefficient determination unit 1102, and the combination of the formant enhancement filter 1003 and the filter coefficient determination unit 1102 included therein is arbitrary.

(Eighth Embodiment) FIG. 15 is a block diagram showing an eighth embodiment. In FIG. 15, FIG.
Components having the same reference numerals as those having the same functions as those in FIG.

The present embodiment is applied to a preprocessing unit for arbitrary audio processing, and aims at enhancing the formant of an audio signal buried in background noise. According to the present embodiment, since the formants of the audio signal are emphasized and the valleys of the audio spectrum are attenuated, the spectrum of the background noise riding on the valleys of the audio spectrum can be attenuated, and the sense of noise can be reduced. Be relaxed.

In FIG. 15, digitized audio signals are sequentially input from an input terminal 2101 and are input to a buffer 2.
At 102, a predetermined number of NF audio signals are input. When a predetermined number of audio signals are input, LP
The audio signal is transferred to C coefficient analysis section 2001 and gain adjustment section 1601. The recommended value of NF is 160. LPC
The coefficient analysis unit 2001 uses the autocorrelation method or the covariance method as described above, but in FIG. 15, analysis is performed by the autocorrelation method. In the autocorrelation method, the LPC coefficient and P
Since the ARCOR coefficient can be obtained at the same time, the formant enhancement filter 1003 uses the LPC coefficient and the PARCO coefficient.
An R coefficient is provided. Similarly, the LPC coefficient analysis unit 2001
It is also possible to use the covariance method for
Only the coefficients are provided to the formant enhancement filter 1003.

Note that in FIG. 15, the configuration of FIG.
02 shows the case where the configuration of FIG. 6 is used,
Using the configuration of FIG. 4 for the formant enhancement filter 1003,
The configuration of FIG. 5 may be used for the filter coefficient determination unit 1102, and the combination of the formant enhancement filter 1003 and the filter coefficient determination unit 1102 included therein is arbitrary.

[Embodiment 2] (Ninth Embodiment) FIG. 16 is a block diagram showing a configuration of a formant enhancement filter according to a ninth embodiment.
In FIG. 16, components denoted by the same reference numerals as in FIG. 2 have the same functions as in FIG.
The present embodiment differs from the present embodiment in the method of realizing the formant enhancement filter 1003. The formant emphasis filter 1003 of the present embodiment includes a polar filter 300
3, a zero-type filter 3004, and these pole-type filters 3
003, a pole-type filter coefficient determination unit 3001 and a zero-type filter coefficient determination unit 3002 that determine respective filter coefficients of the zero-type filter 3004.

The polar filter 3003 emphasizes peaks (formants) of the spectrum of the input audio signal,
Formant emphasis filter 1003 for attenuating valleys
Function as a main filter that achieves the basic function of The zero-type filter 3004 is different from the pole-type filter 30.
03 works to correct the extra spectral tilt. The processing flow of the present embodiment will be described below with reference to FIG.
This will be described with reference to FIG.

The input terminal 1002 sequentially receives LPC coefficients representing the general shape of the spectrum of the audio signal. The polar filter coefficient determination unit 3001 determines the filter coefficient ｛q (i) of the polar filter 3003 based on the input LPC coefficient;
Find i = 1 to P 求 め る. Similarly, zero-type filter coefficient determination section 3002 obtains filter coefficients {r (i); i = 1 to P} of zero-type filter 3004. Pole-type filter coefficient determination unit 3001 and zero-type filter coefficient determination unit 3002
The specific processing method will be described later. The audio signal input from the input terminal 1001 is
003 and a zero-type filter 3004 in this order, and a formant-emphasized signal is output from an output terminal 1004.

When the transfer functions of the pole type filter 3003 and the zero type filter 3004 are represented in the z-transform domain, the following equation (20) is obtained. Here, C (z) represents the z-converted value of the input audio signal, and G (z) represents the z-converted value of the output signal.

[0108]

(Equation 9)

If the equation (20) is expressed in the time domain, the following equation is obtained.
It becomes an expression. Here, c (n) is represented as a time domain signal of C (z), and g (n) is represented as a time domain signal of G (z).

[0110]

(Equation 10)

Next, the pole type filter coefficient determining section 3001 and the zero type filter coefficient determining section 3002 will be described in detail.

FIG. 17 shows a polar filter coefficient determining section 300.
FIG. 9 is a block diagram illustrating a first configuration example of a filter coefficient determination unit that can be applied to the 1 and zero type filter coefficient determination unit 3002. In the figure, each of the LPC coefficients {α _i ; i = 1 to P} inputted from the input terminal 1002 is represented by a constant λ ⁱ (i: the order of the LPC coefficient). Are output from the output terminal 3006 as filter coefficients. For example, when the filter coefficient determination unit having the configuration shown in FIG. 17 is used as the polar filter coefficient determination unit 3001, the filter coefficients {q (i); i = 1 to P} of the polar filter 3003 are represented by q (i) = α _i λ ⁱ (22)

Similarly, zero-type filter coefficient determination section 3002
, The filter coefficient ｛r of the zero-type filter 3004
(I); i = 1 to P｝ is determined as r (i) = α _i λ ⁱ (23).

Next, a description will be given of a second example of the configuration of the filter coefficient determination unit applicable to the polar filter coefficient determination unit 3001 and the zero type filter coefficient determination unit 3002 with reference to FIG. The configuration of FIG. 18 differs from the configuration of FIG. 17 in that a memory table 3007 storing a constant to be multiplied by each coefficient of the LPC coefficient is stored. In the figure, LPC coefficient ｛α _i input from input terminal 1002; i = 1 to
Each of the following coefficients of P｝ is a constant {t (i); i = 1 to P} arbitrarily determined in advance corresponding to each of the following coefficients stored in the memory table 3007 and the multiplier 3005 respectively. The output is multiplied and output from the output terminal 3006 as a filter coefficient. For example, when the filter coefficient determination unit having the configuration of FIG. 18 is used for the polar filter coefficient determination unit 3001, the filter coefficient ｛q
(I); i = 1 to P｝ are determined as follows: q (i) = α _i t (i) (24)

Similarly, zero-type filter coefficient determination section 3002
, The filter coefficient ｛r of the zero-type filter 3004
(I); i = 1 to P｝ is determined as r (i) = α _i t (i) (25).

A feature of this embodiment is that at least one of the polar filter coefficient determining unit 3001 and the zero filter coefficient determining unit 3002 is configured using a memory table 3007 as shown in FIG. . By doing so, the filter coefficient to be multiplied by the LPC coefficient to obtain the filter coefficient can be freely set by the memory table 3007 without being limited to the exponential function value as in the related art. High quality sound can be obtained by the emphasis filter 1003.

The flow of the above processing is summarized as a flowchart shown in FIG. Where ｛c
(N); n = -P ~ NUM-1} represents a signal sequentially input from the input terminal 1001, and {g (n); n = -P ~
NUM-1} represents an output signal. When the variable n of c (n) and g (n) is a negative value, it means that the data of the previous frame is used. Here, NUM represents the frame length, and NUM = 40. Further, P represents a filter order, and here, P = 10. Steps S41, S45, and S46 in FIG. 19 are the same as steps S11, S15, and S16 in FIG. 3 described above, and a description thereof will be omitted.

The processes newly provided in FIG. 19 are steps S42 to S44 and S47. The features of this processing are that filtering is performed using a P-order pole-type filter and a P-order zero-type filter, and a method for calculating filter coefficients of a pole-type filter and a filter coefficient for a zero-type filter and a method for updating a filter internal state. is there. Hereinafter, step S
42 to S44 and S47 will be described.

In step S42, using the LPC coefficient {α _i ; i = 1 to P} representing the general shape of the spectrum of the input voice signal, the filter coefficient of the polar filter {q (i); = 1 to P｝. Step S43
Then, the filter coefficient {r (i); i = 1 to P} of the zero-type filter is calculated according to the equation (25). Next, in step S44, a filtering process of the polar filter and the zero filter is performed according to the equation (21). In step S47, the internal state of the filter is updated according to equations (26) and (27) for the next frame.

C (j-NUM) = c (j) (j = NUM-P to NUM-1) (26) g (j-NUM) = g (j) (j = NUM-P to NUM-1) (27) In the above-described processing, an example has been described in which the filter coefficient of the polar filter is calculated according to equation (22), and the filter coefficient of the zero-type filter is calculated according to equation (25). It is only necessary that at least one filter coefficient of the polar filter and the zero filter is calculated according to the equation (24) or (25). Further, the filtering order in the filtering process in step S44 is arbitrary. However,
When the order is changed, it is natural that it is necessary to allocate an internal state that matches the order.

(Tenth Embodiment) FIG. 20 is a block diagram showing the configuration of a formant emphasis filter 1003 according to a tenth embodiment. A zero-type filter 3004 for correcting a spectral gradient inherent in a polar filter 3003 is shown. 16 in that it has an auxiliary filter 3008 that works to assist the function of FIG. As the auxiliary filter 3008, the above-described fixed characteristic filter 1104 may be used. However, since the purpose of the auxiliary filter 3008 is to assist the function of correcting the spectral tilt of the zero-type filter 3004 as described above, the characteristic does not necessarily have to be fixed. For example, a spectral tilt such as a PARCOR coefficient is expressed. It may be a filter whose characteristics change depending on the possible parameters. The order of each filter is not limited to that shown in FIG.

(Eleventh Embodiment) FIG. 21 is a block diagram showing a configuration of a formant emphasis filter 1003 according to an eleventh embodiment.
Is different from FIG. Also in this case, the order of each filter is not limited to that shown in FIG. 21 and is arbitrary.

(Twelfth Embodiment) FIG. 22 is a block diagram showing a configuration of a formant emphasis filter 1003 according to a twelfth embodiment, which differs from FIG. 16 in having an auxiliary filter 3008 and a pitch emphasis filter 2202. .
Also in this case, the order of the filters is arbitrary.

(Thirteenth Embodiment) FIG. 23 is a block diagram showing a configuration of a formant emphasis filter 1003 according to a thirteenth embodiment. The feature of this embodiment is that the polar filter coefficient determination unit 3001 and the zero-type filter coefficient determination unit 3002 perform M types (M ≧ 2) of constants ｛λ _m ;
M} or memory table {t _m (i); i = 1 to
P, m = 1 to M｝, and one of M types of constants or one of memory tables is selected according to the attribute of the input audio signal and used to determine a filter coefficient.

The operation will be described below by focusing on the features of the present embodiment. However, the pole filter coefficient determination section 3001, (22) the filter coefficients are determined using a constant lambda _m according equation, the zero-type filter coefficient determining section 3002 (25) memory table {t _m according equation (i ); I = 1
The description will be made on the assumption that the filter coefficient is determined by using P｝. Note that at least one of the polar filter coefficient determination unit 3001 and the zero-type filter coefficient determination unit 3002 may be configured to determine a filter coefficient using a memory table according to the equation (24) or (25). Therefore, it is not necessary to be limited to the above configuration.

In FIG. 23, information (attribute information) representing the attribute of the input audio signal is input from input terminal 3009,
It is provided to a pole type filter coefficient determination unit 3001 and a zero type filter coefficient determination unit 3002. The polar filter coefficient determination unit 3001 selects one constant from among M kinds of prepared constants {λ _m ; m = 1 to M} based on the attribute information, and uses the selected λ _m. The coefficient of the polar filter 3003 is calculated according to the equation (22). Similarly, in the zero-type filter coefficient determination unit 3002, based on the attribute information,
M types are prepared and stored in the memory table are constants _{{t m (i); i} = 1~P, m = 1~M} 1 from among
Determining the filter coefficients of using (25) the zero type according formula filter 3004; One of the select memory table, selected constants stored in the memory table _{{i = 1~P t m (i} )}.

The attribute information of the input audio signal is, for example, information indicating the type of a vowel part, a consonant part, and a background part. When such attribute classification is performed, it is effective to set so that formant emphasis is performed strongly in vowel parts, and then formant emphasis is weakened in the order of consonant parts and background parts. As a method of classifying attributes, for example, a primary P
In some cases, a method of classifying using a plurality of types of characteristic parameters such as an ARCOR coefficient and a pitch gain can be used.

FIG. 24 is a diagram showing the polar filter coefficient determining section 3001 and the zero filter coefficient determining section 300 shown in FIG.
FIG. 5 is a block diagram illustrating a first configuration example of a filter coefficient determination unit applied to the second embodiment. First, based on the attribute information input from the input terminal 3009, M kinds of prepared constants ｛λ
_m ; one constant λ _m is selected from m = 1 to M｝. Then, LPC coefficients input from the input terminal 1002; in each order of {α _{i i} = 1~P}, constants λ _m ^{i (i:} L
, And the resulting filter coefficient is output from an output terminal 3006.

FIG. 25 shows pole type filter coefficient determining section 3001 and zero type filter coefficient determining section 300 in FIG.
FIG. 11 is a block diagram illustrating a second configuration example of a filter coefficient determination unit applied to the second embodiment. First, based on the attribute information input from the input terminal 3009, constants Δt _m (i) stored in M types of prepared memory tables 3007, 3010, 3011; i = 1 to P, m = 1 to M } one memory table is selected from among, stored in the memory table constants _{{t m (i); i} = 1~P} are retrieved. Next, the LPC coefficient {α _i ; i = 1 to P} input from the input terminal 1002 is multiplied by a constant t _m (i) extracted from the memory table to obtain a filter coefficient. Output from the output terminal 3006.

The flow of the above processing is summarized as a flowchart shown in FIG. Where ｛c
(N); n = -P ~ NUM-1} represents a signal sequentially input from the input terminal 1001, and {g (n); n = -P ~
NUM-1} represents an output signal. When the variable n of c (n) and g (n) is a negative value, it means that the data of the previous frame is used. Here, NUM represents a frame length, and NUM = 40. Further, P represents a filter order, and here, P = 10. Also,
Steps S51, S54, S55, S5 in FIG.
6, S57, S58, and S59 correspond to S41, S42, S43, S44, S45, and S4 in FIG. 28 described above.
6, the processing is the same as that in S47, and a description thereof will be omitted.

The processing newly provided in FIG. 29 is steps S52 to S53. Feature of this process, in accordance with the attribute information of the input audio signal, M kinds of stored in the memory table constants _{{t m (i); i} = 1~P, m = 1~M}
Step S52 of selecting a constant stored in one memory table from M, and M kinds of constants ｛λ according to the attribute information.
_m ; Step S for selecting one constant from m = 1 to M｝
At 53.

(Fourteenth Embodiment) FIG. 26 is a block diagram showing a configuration of a formant emphasis filter 1003 according to a fourteenth embodiment, in which an auxiliary filter 3008 is added to the configuration of FIG.

(Fifteenth Embodiment) FIG. 27 is a block diagram showing a configuration of a formant emphasis filter 1003 according to a fifteenth embodiment, in which a pitch emphasis filter 2202 is added to the configuration of FIG.

(Sixteenth Embodiment) FIG. 28 is a block diagram showing a configuration of a formant emphasis filter 1003 according to a sixteenth embodiment, in which an auxiliary filter 3008 and a pitch emphasis filter 2202 are added.

In the fourteenth to sixteenth embodiments, the order of each filter can be arbitrarily changed.

(Seventeenth Embodiment) As a seventeenth embodiment, FIG. 30 shows another embodiment in which the present invention is applied to a speech decoding apparatus of a speech encoding / decoding system. FIG.
In FIG. 1, components denoted by the same reference numerals as those in FIG.
The description is omitted assuming that it has the same function as 1.

The difference between the fifth embodiment and the seventeenth embodiment is that the fifth embodiment has a formant enhancement filter based on the configuration shown in FIG.
Embodiment 0 is characterized in that a formant enhancement filter based on the configuration shown in FIG. 16 is applied.

In FIG. 30, polar filter coefficient determination section 3001 outputs LP output from coefficient conversion section 1712.
The LPC coefficient に 従 い α _i ; i
= 1 to P} and a constant λ ⁱ (i: the order of the LPC coefficient) to calculate a polar filter coefficient {q (i); i = 1 to P}. In the zero-type filter coefficient determination unit 3002, (2
According to the expression 5), the product of the LPC coefficient {α _i ; i = 1 to P} and the constant {t (i); i = 1 to P} stored in the memory table 3007 prepared in advance is obtained, Calculate the filter coefficient {r (i); i = 1 to P}.

The synthesized signal output from the synthesis filter 1709 is output from the pitch emphasis filter 22 expressed by the equation (16).
02, the pitch is emphasized. As the pitch period L at this time, a pitch period calculated from the adaptive codebook index IACB is used. The pitch filter gain uses a predetermined fixed value δ (for example, δ = 0.7). In this embodiment, pitch enhancement is performed using the pitch period calculated from the adaptive codebook index IACB. However, the present invention is not limited to this. For example, the output signal of the synthesis filter 1709 or the output signal of the adder 1708 may be used. A new analysis may be performed to determine the pitch period. Further, the pitch filter gain is not limited to a fixed value, and for example, a method of calculating from the output signal of the synthesis filter 1709 or the output signal of the adder 1708 may be used.

Next, formant enhancement is performed through the polar filter 3003, the zero filter 3004, and the auxiliary filter 3008. Here, as the auxiliary filter 3008, a fixed characteristic filter represented by the equation (9) is used. Formant enhancement filter 1003 in gain adjustment section 1601
Is processed so that the output signal power becomes equal to the input signal power and the change in power becomes smooth, and is output as a final synthesized speech signal.

The order of each filter is not limited to the order described above, but is arbitrary. Further, in the present embodiment, the case has been described where the pitch enhancement filter 2202 and the auxiliary filter 3008 exist as components of the formant enhancement filter 1003. However, the configuration excluding the pitch filter 2202, the configuration excluding the auxiliary filter 3008, and both of them are described. The configuration may be omitted. In the present embodiment, the polar filter coefficient determination unit 3001 uses a coefficient determination method according to the equation (22), and the zero-type filter coefficient determination unit 3002 uses (2
The case where the coefficient determination method according to the expression 5) is used has been described. However, the present invention is not limited to this. At least one of the polar filter coefficient determination unit 3001 and the zero-type filter coefficient determination unit 3002 uses the expression (24). Alternatively, a coefficient determination method according to the equation (25) may be used.

(Eighteenth Embodiment) As an eighteenth embodiment, FIG. 31 shows another embodiment in which the present invention is applied to a speech decoding apparatus of a speech encoding / decoding system. FIG.
In FIG. 3, the components denoted by the same reference numerals as those in FIG.
The description is omitted because it has the same function as 0.

The difference between the seventeenth embodiment and the eighteenth embodiment is that in the seventeenth embodiment, the formant emphasis filter 1
003 regardless of the attribute of the input audio signal, the fixed value λ of the polar filter coefficient determination unit 3001 and the value Δt (i) stored in the memory table 3007 of the zero-type filter coefficient determination unit 3002; i = 1 to P ｝ Always has the same value, whereas in the eighteenth embodiment, M types {λ _m ; m = 1 to M} prepared according to the attributes of the input audio signal and stored in the memory tables 3007, 3010, 3011. constant _{{t m (i); i} = 1~P, m = 1~M} by selecting one each from among, except that calculates the filter coefficients.

FIG. 31 shows a fixed value {λ _m ; m = 1 to M} and a constant {t stored in the memory table 3007.
_m (i); When selecting i = 1 to P, m = 1 to M｝,
A configuration in the case where an attribute of an input audio signal is transmitted as additional information from an encoder (not shown) is shown. Therefore, the attribute information is decoded by the demultiplexer 1702, and a fixed value and a memory table are selected based on the information.

In this embodiment, the case where the attribute information of the input audio signal is transmitted from the encoder unit has been described. However, the spectrum information and the adaptive gain of the adaptive gain obtained from the decoded LPC coefficient without using such additional information. The attribute may be determined based on the decoding parameter such as the size. in this case,
Since no additional information is required, an increase in the transmission rate can be avoided.

(Nineteenth Embodiment) As a nineteenth embodiment, FIG. 32 shows another embodiment in which the present invention is applied to a speech decoding apparatus of a speech encoding / decoding system. FIG.
In FIG. 3, the components denoted by the same reference numerals as those in FIG.
The description is omitted because it has the same function as 0.

The difference between the seventeenth embodiment and the nineteenth embodiment is that, in the seventeenth embodiment, the polar filter coefficient and the zero filter coefficient are calculated based on the decoded LPC coefficients. In the embodiment, the LPC coefficient analysis is performed again on the synthesized signal obtained from the synthesis filter 1709, and the polar filter coefficient and the zero-type filter coefficient are calculated based on the LPC coefficients obtained as a result.
By doing so, formant emphasis can be performed accurately as described in the seventh embodiment.
Since the analysis order of the PC coefficient can be arbitrarily set, the formant enhancement can be more finely controlled when the analysis order is increased.

(Twentieth Embodiment) As a twentieth embodiment, FIG. 33 shows another embodiment in which the present invention is applied to a speech decoding device of a speech encoding / decoding system. FIG.
In FIG. 3, the components denoted by the same reference numerals as those in FIG.
The description is omitted assuming that it has the same function as 1.

The difference between the nineteenth embodiment and the twentieth embodiment is that, in the nineteenth embodiment, the polar filter coefficient and the zero filter coefficient are calculated based on the decoded LPC coefficients. In the embodiment, the LPC coefficient analysis is performed again on the synthesized signal obtained from the synthesis filter 1709, and the polar filter coefficient and the zero-type filter coefficient are calculated based on the LPC coefficients obtained as a result.
By doing so, formant emphasis can be performed accurately as described in the seventh embodiment.
Since the analysis order of the PC coefficient can be arbitrarily set, the formant enhancement can be more finely controlled when the analysis order is increased.

(Twenty-First Embodiment) As a twenty-first embodiment, FIG. 34 shows another embodiment in which the present invention is applied to a preprocessing section for an arbitrary voice processing. 34, components having the same reference numerals as those in FIGS. 15 and 32 have the same functions as those in FIGS. 15 and 32, and a description thereof will be omitted.

The difference between the eighth embodiment and the twenty-first embodiment is that the eighth embodiment has a formant emphasis filter based on the configuration shown in FIG.
In the first embodiment, a formant enhancement filter based on the configuration shown in FIG. 16 is applied.

(Twenty-second Embodiment) As a twenty-second embodiment, FIG. 35 shows another embodiment in which the present invention is applied to a pre-processing unit for arbitrary sound processing. 35, components having the same reference numerals as those in FIG. 34 have the same functions as those in FIG.

The difference between the twenty-first embodiment and the twenty-second embodiment is that, in the twenty-first embodiment, regardless of the attribute of the input audio signal, the fixed value λ and the zero-type filter coefficient Memory table 3 of unit 3002
The constants {t (i); i = 1 to P} stored in 007 are always the same, whereas in the twenty-second embodiment, M kinds of constants {λ _m ; m =
1 to M} and memory tables 3007, 3010, 3
Constant １t _m (i) stored in 011; i = 1 to P, m
= 1 to M}, and the filter coefficient is calculated.

FIG. 35 shows fixed values {λ _m ; m = 1 to M} and constants {t _m (i) stored in memory tables 3007, 3010, 3011; i = 1 to P, m = 1 to M ｝
Is selected, the input audio signal stored in the buffer 2102 and the LP output from the LPC coefficient analysis unit 2001 are selected.
Attribute classification unit 301 using C coefficient {α _i ; i = 1 to P}
At 3, the attributes of the input audio signal are analyzed. Then, based on the analysis result, M kinds of constants {λ _m ; m = 1 to M} and constants {t _m (i) stored in the memory tables 3007, 3010, 3011; i = 1 to P, m = 1 ｝ M} is selected for use in the frame, and used for filter coefficient calculation. The attribute classification unit 3013 determines an attribute using the spectrum information and the pitch information of the input audio signal.

[0155]

As described above, according to the present invention, it is possible to provide a formant emphasizing method capable of obtaining high-quality sound.

That is, in the first aspect, the formant enhancement processing for enhancing the peaks of the spectrum of the input audio signal and attenuating the valleys is performed, and the spectral inclination caused by the formant enhancement is determined by the characteristic of the input audio signal. Alternatively, the formant enhancement of the audio signal and the correction of the extra spectral tilt accompanying the small amount of processing to be corrected by the primary filter whose characteristic is adaptively changed according to the spectrum enhancement characteristic and the primary filter whose characteristic is fixed are performed. It can be performed effectively, and the subjective quality can be greatly improved.

In the second invention, a formant enhancement process for enhancing peaks and attenuating valleys of a spectrum of an input audio signal is performed by a polar filter, and a spectrum inclination caused by the formant enhancement is corrected. The processing is performed by a zero-type filter, and at least one of the filter coefficient of the polar filter and the filter coefficient of the zero-type filter is made to correspond to each coefficient of the LPC coefficient of the input audio signal and each coefficient of the LPC coefficient. By determining by a product with a constant arbitrarily determined in advance, the filter coefficient of the formant enhancement filter can be finely controlled,
As a result, high-quality sound can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining a basic operation of a formant enhancement filter according to a first embodiment;

FIG. 2 is a block diagram showing a configuration of a formant enhancement filter according to the embodiment;

FIG. 3 is a flowchart showing a processing procedure of the embodiment.

FIG. 4 is a block diagram showing a configuration of a formant enhancement filter according to a second embodiment.

FIG. 5 is a block diagram illustrating a configuration example of a filter coefficient determination unit according to the first and second embodiments.

FIG. 6 is a flowchart showing a processing procedure when the filter coefficient determination unit of FIG. 5 is used.

FIG. 7 is a block diagram showing another configuration example of the filter coefficient determination unit in the first and second embodiments.

8 is a flowchart showing a processing procedure when the filter coefficient determination unit of FIG. 7 is used.

FIG. 9 is a block diagram illustrating a configuration of a formant enhancement filter according to a third embodiment.

FIG. 10 is a block diagram illustrating a configuration of a speech decoding apparatus according to a fourth embodiment.

FIG. 11 is a block diagram showing a configuration of a speech decoding apparatus according to a fifth embodiment.

FIG. 12 is a block diagram showing a configuration of a speech decoding apparatus according to a sixth embodiment.

FIG. 13 is a block diagram showing a basic operation of the formant emphasis filter in the embodiment.

FIG. 14 is a block diagram showing a configuration of a speech decoding apparatus according to a seventh embodiment.

FIG. 15 is a block diagram showing a configuration of an audio preprocessing device according to an eighth embodiment.

FIG. 16 is a block diagram showing a configuration of a formant enhancement filter according to a ninth embodiment;

FIG. 17 is a block diagram showing a configuration example of a filter coefficient determination unit according to the embodiment;

FIG. 18 is a block diagram showing another configuration example of the filter coefficient determination unit in the embodiment.

FIG. 19 is a flowchart showing a processing procedure according to the embodiment;

FIG. 20 is a block diagram showing a configuration of a formant enhancement filter according to a tenth embodiment.

FIG. 21 is a block diagram showing a configuration of a formant enhancement filter according to an eleventh embodiment.

FIG. 22 is a block diagram showing a configuration of a formant enhancement filter according to a twelfth embodiment.

FIG. 23 is a block diagram showing a configuration of a formant enhancement filter according to a thirteenth embodiment.

FIG. 24 is a block diagram showing a configuration example of a filter coefficient determination unit in the embodiment.

FIG. 25 is a block diagram showing another configuration example of the filter coefficient determination unit in the embodiment.

FIG. 26 is a block diagram showing a configuration of a formant enhancement filter according to a fourteenth embodiment.

FIG. 27 is a block diagram showing a configuration of a formant enhancement filter according to a fifteenth embodiment.

FIG. 28 is a block diagram showing a configuration of a formant enhancement filter according to a sixteenth embodiment.

FIG. 29 is a flowchart showing the processing procedure of the thirteenth to sixteenth embodiments.

FIG. 30 is a block diagram showing a configuration of a speech decoding device according to a seventeenth embodiment.

FIG. 31 is a block diagram showing a configuration of a speech decoding apparatus according to an eighteenth embodiment.

FIG. 32 is a block diagram showing a configuration of a speech decoding apparatus according to a nineteenth embodiment.

FIG. 33 is a block diagram showing a configuration of a speech decoding apparatus according to a twentieth embodiment.

FIG. 34 is a block diagram showing a configuration of an audio preprocessing device according to a twenty-first embodiment.

FIG. 35 is a block diagram showing a configuration of an audio preprocessing device according to a twenty-second embodiment.

[Explanation of symbols]

1001 ... Audio signal input terminal 1002 ... LPC coefficient input terminal 1003 ... Formant emphasis filter 1004 ... Audio signal output terminal 1101 ... Spectral emphasis filter 1102 ... Filter coefficient determination unit 1103 ... Primary characteristic variable filter 1104 ... Primary characteristic fixed filter 1401 ... Filter coefficient conversion unit from LPC coefficient to PARCOR coefficient 1402,1706,1707 ... Multiplier 1501 ... Filter coefficient restriction unit 1502 ... Buffer 1601 ... Gain adjustment unit 1712 ... Coefficient conversion unit from LSP coefficient to LPC coefficient 1702 ... De Multiplexer 1703 LSP coefficient inverse quantizer 1704 Adaptive codebook 1705 Noise codebook 1708 Adder 1709 Synthesis filter 1710 Adaptive gain decoder 1711 Noise gain decoder 18 01: PARCOR coefficient inverse quantization unit 1802: Coefficient conversion unit from PARCOR coefficient to LPC coefficient 2001: LPC coefficient analysis unit 2102: Buffer 2201: Input terminal of pitch period and pitch gain 2202: Pitch emphasis filter 3001: Polar filter coefficient Determining unit 3002: zero-type filter coefficient determining unit 3003: polar filter 3004: zero-type filter 3005, 3012: multiplier 3007, 3010, 3011: memory table (constant storage means) 3008: auxiliary filter

Continuing from the front page (72) Inventor Akinobu Yamashita 1 Toshiba R & D Center, Komukai, Kawasaki-shi, Kanagawa Prefecture (56) References JP-A-64-13200 (JP, A) 2-82710 (JP, A) JP-A-7-160296 (JP, A) JP-A-8-6596 (JP, A) JP-A-63-8700 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/06 G10L 19/02

Claims (3)

(57) [Claims] 1. A formant emphasis method for emphasizing peaks of a spectrum of an input audio signal and attenuating valleys, said formant emphasis emphasizing peaks of a spectrum of an input audio signal and attenuating valleys. In addition to performing the processing, the spectral gradient caused by this formant enhancement processing is corrected by a primary filter whose characteristics are adaptively changed according to the characteristics of the input audio signal or the spectral enhancement characteristics and a primary filter whose characteristics are fixed. A formant enhancement method characterized by the following. 2. A formant emphasis method for emphasizing peaks of a spectrum of an input audio signal and attenuating valleys, said formant emphasis emphasizing peaks of a spectrum of an input audio signal and attenuating valleys. The processing is performed by a polar filter, and the processing for correcting the spectrum tilt caused by the formant enhancement processing is performed by a zero filter. At least one of the filter coefficient of the polar filter and the filter coefficient of the zero filter is converted to the input voice. A formant emphasizing method characterized in that it is determined by a product of a coefficient arbitrarily determined in advance corresponding to each coefficient of the LPC coefficient of the signal and each coefficient of the LPC coefficient of the signal. 3. A formant emphasis filter for emphasizing a peak portion of a spectrum of an input audio signal and attenuating a valley portion, wherein the formant emphasis filter emphasizes a peak portion of the spectrum of the input audio signal and attenuates a valley portion. The process is performed by a polar filter, the process of correcting the spectrum tilt caused by this formant enhancement process is performed by a zero-type filter, and a plurality of constants arbitrarily determined in advance corresponding to the respective following coefficients of the LPC coefficient. A plurality of stored constant storage means are provided, and at least one of a filter coefficient of the polar filter and a filter coefficient of the zero filter is a coefficient of each order of LPC coefficients of the input audio signal, and an attribute of the input audio signal. The corresponding constant stored in one constant storage means selected from the plurality of types of constant storage means based on Formant emphasis method characterized by determining in the product.