JP3426871B2 - Method and apparatus for adjusting spectrum shape of audio signal - Google Patents

Abstract

PROBLEM TO BE SOLVED: To stably improve the quality of voices such as decoded voice, synthesized voice and so forth with a small amt. of calculation. SOLUTION: This device cascade connects a first filter 112 having pole-zero type transfer functions A(z)/B(z) for emphasizing the spectral envelope of a voice signal and a second filter 113 for compensating inclinations of spectrums of the voice signal. Then, two filter coefficients to be used in the second filter 113 in order to compensate the inclinations of the spectrums are respectively calculated independently from the pole-zero type transfer functions and then the inclinations of spectrums respectively corresponding to pole-zero transfer functions are compensated by the calculated filter coefficients.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for adjusting the spectral shape of a speech signal in order to improve the quality of decoded speech and synthesized speech.

[0002]

2. Description of the Related Art An audio signal is encoded at a low bit rate,
In a voice encoding / decoding system in which the obtained encoded data is passed through a transmission system or a storage system and then the encoded data is decoded, in order to improve the subjective quality of the voice signal decoded and reproduced on the decoding side. In many cases, a post filter is arranged at the final stage of the decoding device.

In a conventional speech decoding apparatus combined with a post filter, various parameters included in encoded data are decoded by a parameter decoding section, and a speech signal reproducing section produces a speech signal based on the decoded parameter information. Is played.

A post filter is arranged at the subsequent stage of the decoding device including the parameter decoding unit and the audio signal reproducing unit. The post filter is configured by connecting a pitch component emphasis filter, a spectrum envelope emphasis filter, a high frequency emphasis filter, and a gain adjustment section in cascade.

The functions of the post filter are roughly classified into emphasis of pitch component, emphasis of spectrum envelope, emphasis of high frequency component, and adjustment of filter gain. Of these, the pitch component and the spectral envelope are important factors that determine the pitch and the phoneme of the voice, and by emphasizing these, the effect of producing a clearer and less noisy voice is obtained. Further, the adjustment of the filter gain is necessary so that the size of the audio signal does not change between the input and the output of the post filter.

The enhancement of high frequency components is carried out in order to compensate for the insufficiency of the high frequency components of the voice caused by the characteristics of the coding and the post filter, such as "comfortable feeling" and "bad sound passage". In particular, a filter used for spectral envelope emphasis often has an unnecessary spectrum slope (on average, a low-frequency emphasis slope), and high-frequency component emphasis is used for the purpose of compensating for this.

[0007]

In the prior art, for example, C (z) = 1-μz ^-1 (μ =
A filter having a fixed transfer function of (a fixed value of about 0.4) is used. When such a high-frequency emphasis filter is used, the "muffled feeling" is in the low frequency range, and the subjective sound quality is improved to some extent. On the other hand, on the other hand, the sound in a section where high-frequency emphasis does not need to be emphasized, such as a consonant section, is excessively emphasized in the high-frequency range, which may cause abnormal noise in the high-frequency range. There is a problem that a sufficient sound quality improvement effect cannot be obtained.

That is, when listening to and analyzing a sound with a muffled sound, the sound is not always muffled, and since the time of the sound section in which the high frequency is not sufficient is long, Sounds like you're muddy. Also,
Depending on the section of sound, the extent to which the high range is not produced also varies.
For this reason, when a high-frequency emphasis filter having a fixed transfer function is used, high-frequency emphasis is performed even in a section in which a high frequency is relatively high, so that sound quality is deteriorated.

As another conventional technique, a method of predictively analyzing the transfer function F (z) of the spectral envelope emphasis filter and adaptively changing the value of the parameter μ in the transfer function C (z) of the high band emphasis filter based on the predictive analysis. It has been known. However, in this method, since the transfer function F (z) of the spectral envelope emphasis filter is generally expressed by that of a pole-zero filter having many orders, there is a problem that calculation for obtaining the parameter μ becomes very complicated. There is.

As described above, in the conventional post filter using the high-frequency emphasis filter whose transfer function is fixed, the difference in the high-frequency range is caused by the excessive high-frequency emphasis of the voice in the section where the high-frequency emphasis is unnecessary. There is a problem that sound is generated, and the post filter that predicts the transfer function of the spectral envelope emphasis filter and adaptively changes the transfer function of the high-frequency emphasis filter based on it, requires a large amount of calculation. There was a problem that it was difficult to implement.

[0011]

SUMMARY OF THE INVENTION It is an object of the present invention to provide a method and apparatus for adjusting the spectral shape of a speech signal, which can stably improve the quality of speech such as decoded speech or synthesized speech with a small amount of calculation. Especially.

Another object of the present invention is to provide a method for adjusting the spectrum shape of a voice signal, which can avoid the deterioration of the sound quality when the gain adjustment is performed along with the adjustment of the spectrum shape of the voice signal.

According to the present invention, a first filter having a pole-zero type transfer function represented by A (z) / B (z) for an audio signal, and a second filter for compensating for the characteristic of the first filter are provided. In the method of adjusting the spectrum shape of the audio signal by passing through the filter (1), the two parameters used for the second filter are obtained separately from A (z) and B (z).

Further, the present invention provides a spectrum shape adjusting device for adjusting the spectrum shape of an audio signal, wherein A (z)
/ B (z) is represented by a first filter having a pole-zero type transfer function, and first and second parameters cascade-connected to the first filter for compensating the characteristics of the first filter. And a parameter generating means for separately generating the first and second parameters from A (z) and B (z).

Here, more specifically, the parameter generation means predicts the characteristic of A (z) or 1 / A (z), and generates the prediction coefficient as the first parameter. And a second parameter generating means for predicting the characteristic of B (z) or 1 / B (z) and generating the prediction coefficient as the second parameter.

The transfer function F (z) = A of the first filter (spectral envelope emphasis filter) of the pole-zero type.
The slope of the spectrum of (z) / B (z) is A on the molecular side.
(Z) and the denominator side B (z) can be expressed by combining the slopes of the respective spectra. The present invention pays attention to this point, separates F (z) into A (z) and B (z), and from these filter coefficients of A (z) and B (z),
The two parameters of the second filter (adaptive filter) are directly obtained. The adaptive filter is configured to have a transfer function that compensates for the slope of the spectrum of the entire F (z) by combining the two parameters thus obtained.

In this way, the zero-type shape of F (z) having a small order is formed without predicting the overall characteristics of F (z), which is a pole-zero type having a large order, as in the prior art. From the information of the filter coefficients of A (z) and B (z) expressed by, the parameters necessary for the adaptive filter that can compensate the slope of the spectrum of F (z) easily and accurately can be obtained. That is, it is possible to stably improve the sound quality with a small amount of calculation.

Further, according to the present invention, the second filter is characterized by performing at least a filtering process using a polar filter and a zero filter having different parameters. Here, it is desirable that both the pole-type filter and the zero-type filter are low-order filters in order to reduce the amount of calculation, and in terms of the amount of calculation, the first-order filter is optimal. When implementing the second filter at the primary filter, the second filter has a transfer function of z-transform domain ^{_{(1-μ z z -1)}} / (1-μ p z -1) ( where, mu _z , Μ _p are realized by a filter including at least a first-order pole-zero type transfer function represented by mutually independent filter coefficients each having an absolute value smaller than 1.

In this way, the second filter is used as a pole first-order zero
When configured with the following filter, the conventional filter configured with a first-order zero-type filter has one parameter, but has two parameters, so that the transfer function of the second filter is The degree of freedom of expression is increased, the inclination of the spectrum can be compensated flexibly, and the effect of improving the sound quality is further enhanced.

[0020] In this case, we seek mu _p from A (z), if the B (z) to determine the mu _z, it is possible to compensate for the tilt of the spectrum at lower filter coefficients of orders.

Further, in this case, using the weighting coefficient of the relationship of C ₁ <C ₃ <C ₀ , the first autocorrelation coefficient obtained from the parameter of A (z) has a threshold value (Th and Do)
When it is smaller, the first autocorrelation coefficient is weighted with the weighting coefficient C ₀ , and when it is larger than the threshold Th, the first autocorrelation coefficient is weighted with the weighting coefficient c 1. Μ _p is obtained from the value obtained by performing, and the weighting factor C is added to the second autocorrelation coefficient obtained from the parameter of B (z).
If μ _z is obtained from the value obtained by weighting with ₃ , if the first autocorrelation coefficient is smaller than the threshold Th, the voice in this section is a voice like a strong consonant in the high frequency range. On the contrary, when the first autocorrelation coefficient is larger than the threshold value Th, the voice in this section is a voice like a strong vowel in the low frequency range. By switching the weighting factor by comparing with the threshold value Th, the second filter can have a compensation characteristic suitable for each consonant and vowel, and the sound quality is effectively improved.

Furthermore, the present invention is a method for adjusting the spectral shape of an audio signal by performing a predetermined filtering process on the audio signal, in order to compensate for the power change of the audio signal due to the filtering process. When adjusting the gain to be multiplied, the positive / negative of the gain to be multiplied by the audio signal is determined, and if the gain is negative, the gain is replaced with a non-negative value given by a predetermined method. Here, the non-negative value to be replaced is preferably a small value, specifically, a value of 0 or more and less than 1. In this way, it is possible to avoid the sound quality deterioration due to the negative gain when performing the gain adjustment accompanying the adjustment of the spectrum shape of the audio signal.

According to the present invention, a first filter having a pole-zero type transfer function for emphasizing a spectrum envelope of a voice signal and a slope of the spectrum of the voice signal by the filtering process performed by the first filter are compensated. A second filter for adjusting the spectrum, and a step of independently obtaining two filter coefficients used in the second filter for compensating for the slope of the spectrum from the pole-zero type transfer function. And a step of compensating the slope of the spectrum corresponding to each of the pole-zero type transfer functions by the obtained filter coefficient.

According to the present invention, the first filter having a pole-zero type transfer function for enhancing the spectral envelope of the voice signal and the slope of the spectrum of the voice signal by the filtering process performed by the first filter are compensated. Therefore, the audio signal output from the first filter is filtered according to the calculation unit that independently obtains two filter coefficients from the pole-zero type transfer function that is input from the first filter and the obtained filter coefficient. And a second filter having a filter section for compensating for the slope of the spectrum corresponding to each of the pole-zero type transfer functions, and a spectrum shape adjusting device for an audio signal are provided.

According to the present invention, a synthesis filter analyzer that analyzes an input voice signal and outputs synthesis filter data, and a filter data calculator that calculates weighted filter data based on the synthesis filter data from the synthesis filter analyzer. A weighting filter for filtering the input audio signal based on the weighting filter data, the weighting filter compensating for the slope of the spelltle by the first filter having a pole-zero transfer function using the weighting filter data. An apparatus for spectral shape adjustment of an audio signal is provided that includes a second filter having a transfer function for

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION A speech decoding apparatus incorporating a post filter according to a first embodiment of the present invention will be described with reference to FIG. This audio decoding device is composed of a parameter decoding unit 101, an audio signal reproducing unit 102, and a post filter 103.

To the input terminal 100, the encoded data transmitted from the voice encoding device on the transmitting side is input. This encoded data is input to the parameter decoding unit 101, and parameter information such as the pitch vector, noise vector, gain and LPC coefficient used in the audio signal reproducing unit 102 is decoded. The audio signal reproducing unit 102 reproduces an audio signal based on the input parameter information.

As an example of the audio signal reproducing section 102, C
An example is an ELP (Code Excited Linear Prediction) audio signal reproducing unit. In the audio signal reproducing section of this method, the reproduced pitch vector and the noise vector are multiplied by the reproduced gain and then combined to generate a drive signal for the LPC synthesis filter, and the drive signal is passed to the LPC synthesis filter. By doing so, the audio signal is reproduced.

The post filter 103 is arranged at the final stage of the speech decoding apparatus and is used to improve the subjective quality of the reproduced speech signal. The post filter of this embodiment is configured by connecting a pitch component emphasis filter 111, a spectrum envelope emphasis filter 112, a compensation filter 113 and a gain adjusting unit 114 in cascade connection. The compensation filter 113 includes an adaptive filter 121 and a filter coefficient calculation unit 122 that calculates the filter coefficient thereof, and the filter coefficient calculation unit 122 further includes the first parameter calculation unit 1
23 and the second parameter calculation unit 124.
The gain adjusting unit 114 smoothly adjusts the gain so that the audio signal after processing by the post filter 103 has the same power as the audio signal before processing, and outputs the adjusted audio signal to the audio signal output terminal 104. To do.

The post filter 103 will be described in more detail below. The pitch component emphasis filter 111 is
It is a filter used for the purpose of emphasizing the repetition of the pitch period of the audio signal. Pitch component emphasis filter 111
Although various design methods using the pitch period and the pitch gain as parameters can be considered as the design method of, the P (z) = 1 / (1-εβz ^−T ) can be used as an example of the transfer function thereof. Here, T is a pitch period.
Further, β is a pitch gain, ε is a parameter for adjusting the degree of pitch enhancement, and these are set in a relationship of 0 <εβ <1.

The spectral envelope emphasis filter 112 is used for the purpose of emphasizing the shape of the spectral envelope of the audio signal, and its transfer function is F (z). In the CELP method, a method of emphasizing a spectrum envelope by using a pole-zero type filter having a transfer function F (z) shown below as a spectrum envelope emphasizing filter is generally used.

F (z) = A (z) / B (z) (1) where A (z) = 1 / H (z / γ ₁ ) B (z) = 1 / H (z / γ ₂ ) , 0 <γ ₁ <γ ₂ H (z): a filter transfer function representing the spectrum envelope of the audio signal. When such a spectrum envelope enhancement filter 112 is used, the unevenness of the spectrum envelope can be enhanced, and thus the filter passes the post filter 103. The subsequent audio signal is perceived as having a less audible noise. However, with this configuration, various spectrum slopes are added according to changes in the transfer function F (z) that is determined for each voice.

That is, the transfer function F (z) of the spectrum envelope emphasis filter 112 composed of the pole-zero type filter may have a low-frequency emphasis type inclination which cannot be ignored in the entire spectrum. The high-pass emphasis filter of the transfer function C (z) used in the prior art post filter is
In addition to the role of raising the high frequency components that have been degraded by encoding,
The spectrum envelope emphasis filter had a role of roughly compensating for the unnecessary slope of the spectrum of the low-frequency emphasis.

However, the spectral envelope enhancement filter 11
The transfer function F (z) of 2 is set so as to change according to the characteristics of the spectral envelope of the audio signal to be processed.
The slope of the spectrum changes with time. That is, F (z) at a certain moment has low-frequency emphasis characteristics, but F (z) at another moment (for example, a voice section of a consonant) may have high-frequency emphasis. In this case, if a high-frequency emphasis filter having a fixed transfer function C (z) is used as in the conventional case, the high-frequency band of voice is excessively emphasized, which causes abnormal noise.

On the other hand, in this embodiment, the adaptive filter 1
21 and the filter coefficient calculation unit 122 compensates the inclination of the spectrum generated by using the spectrum envelope emphasis filter 112 of the transfer function F (z) represented by the equation (1), and further, if necessary. It is possible to make adjustments such as giving brightness to the sound quality. The parameter calculation units 123 and 23 of the filter coefficient calculation unit 122 input the zero filter transfer function A (z) and the pole filter transfer function B (z) to the adaptive filter 121, respectively, and input two parameters used by the adaptive filter 121. calculate.

Next, the compensation filter 113 will be described in detail. The transfer function F (z) of the spectral envelope emphasis filter 112 shown in Expression (1) is F (z) = A (z) /
B (z), which can be represented separately in pole-type and zero-type filters. Here, A (z) = Σa _i z ⁻ⁱ , a ₀ = 1 and (i = 0 to 10) (2) B (z) = Σb _i z ⁻ⁱ and b ₀ = 1 and (i = 0 to 0) 10) (3)

In the filter coefficient calculation unit 122, the parameter calculation unit 123 regards the filter coefficient of the transfer function A (z) as the impulse response of the transfer function A (z), and the first-order normalized autocorrelation coefficient of this impulse response. The first parameter ρ _A corresponding to
Pass to. Similarly, the coefficient calculation unit 124 uses the transfer function B (z).
Is regarded as the impulse response of the transfer function B (z), and the second parameter ρ _B corresponding to the first-order normalized autocorrelation function of this impulse response is used as the adaptive filter 12
Pass to 1. These parameters ρ _A and ρ _B can be defined by the following equations.

Ρ _A = (Σa _i a _i-1 ) / (Σa _i ² ) (4) ρ _B = (Σb _i b _i-1 ) / (Σb _i ² ) (5) These parameters ρ _A , ρ The value of _B is a first-order prediction coefficient for the impulse response of the filter having the transfer functions A (z) and B (z), respectively. Based on equations (6) and (7) using these parameters ρ _A and ρ _B ,
(Z) and b (z) are determined.

_A (z) = 1−τ _A (ρ _A ) z ⁻¹ (6) b (z) = 1−τ _B (ρ _B ) z ⁻¹ (7) These a (z) and b (z ) Using adaptive filter 1
The transfer function of 21 is set as in equation (8).

D (z) = a (z) / b (z) (8) Here, τ _A () and τ _B () are functions that adjust the values of the parameters ρ _A and ρ _B. By doing so, the slope of the spectrum of the transfer function F (z) by the spectral envelope emphasis filter 112 can be effectively compensated by the adaptive filter 121 of the transfer function D (z).

The transfer function of the equation (8) is (1-μ _z z ^-1 )
It is a first-order pole-zero type transfer function represented by / (1-μ _p z ^-1 ). However, μ _z and μ _p are independent filter coefficients having absolute values smaller than 1, and in this example, μ _z = τ _A (ρ _A ), μ _p = τ _B (ρ _B ).
Is. That is, these transfer functions μ _z and μ _p are transfer functions A
It can be set independently depending on (z) and B (z).

Next, the flow of processing in the above-mentioned post filter 103 will be described using the flowchart shown in FIG. First, the parameter (filter coefficient) of the transfer function F (z) of the spectrum envelope emphasis filter 112 is acquired (step S11). Next, from these parameters, F (z) is separated into a numerator-side transfer function A (z) and a denominator-side transfer function B (z), and given to the parameter calculation units 123 and 124 of the filter coefficient calculation unit 113, respectively. (Step S12).

In the parameter calculation units 123 and 23, the filter coefficients of the transfer functions A (z) and B (z) are transferred to the transfer function A.
(Z) and B (z) are regarded as impulse responses, and parameters ρ _A and ρ _B corresponding to the first-order normalized autocorrelation coefficient of the impulse responses are calculated by equations (4) and (5),
These are passed to the adaptive filter 121. Adaptive filter 12
1, the first-order filters a (z) and b (z) are obtained from the parameters ρ _A and ρ _B by the equations (6) and (7), and the transfer function D (z) shown in the equation (8) is obtained. Is set (step S13). The adaptive filter 121 is a filter a
(Z), b each independently compensate Shinano the slope of the transfer function pole / zero filter in accordance with (z) is I'll be performed et filtering, spectral envelope emphasis filter 11
The process of compensating the inclination of the spectrum in 2 is performed.

Next, the second embodiment will be described. In this embodiment, the external structure is similar to that of the first embodiment shown in FIG. 1, but the design method of the compensation filter 113 is different.

In the first embodiment, the filter slope of the spectrum due to the transfer function A (z) on the numerator side of the transfer function F (z) of the spectral envelope emphasis filter 112 is adjusted by the adaptive filter 1.
21 is compensated by a (z) on the numerator side of the transfer function D (z),
The slope of the spectrum due to the transfer function B (z) on the denominator side of F (z) is compensated by b (z) on the denominator side. On the other hand, in the second embodiment, the spectral tilt due to the transfer function A (z) on the zero side of the transfer function F (z) is compensated for by the filter b (z) on the pole side of D (z) to obtain the transfer function F (z). The spectral tilt due to the transfer function B (z) on the pole side of (z) is compensated by a (z) on the zero point side of D (z). In other words, the transfer function A
Determined coefficient mu _p from (z), so that obtaining the coefficient mu _z transfer function B (z). This is based on the idea that the zero point is compensated by the pole and the pole point is compensated by the zero point, because the compensation can be performed by the filter coefficient having a lower order, and the efficiency is good.

Specifically, by regarding the filter coefficient of the transfer function A (z) as an LPC coefficient, a first-order approximation of the spectral envelope of the transfer function A (z) is performed by using the inverse algorithm of the Durbin method. The PARCOR coefficient (partial autocorrelation coefficient) k _{A of} is obtained as the first parameter of the adaptive filter 121. Similarly, a first-order PARCOR coefficient k _B that approximates the spectral envelope of B (z) is obtained as the second parameter of the adaptive filter 121. At this time,
The parameters k _A and k _B are 1 / A (z) and 1 respectively.
It is a first-order prediction coefficient for the impulse response of / B (z).

Using these two parameters k _A and k _B , the transfer function D of the adaptive filter 121 is adjusted so as to compensate for the spectral tilt caused by A (z) and B (z).
(Z) is set. Its specific example is, D (z) = a ( z) / b (z) (9) a (z) = 1-η B (k B) z -1 (10) b (z) = 1- η _A (k _A ) z ⁻¹ (11). Here, η _A () and η _B () are functions that adjust the values of the parameters k _A and k _B. As an example, η _A (k _A ) = 0.5 k _A , η _B (k _B ) = 0.8
It can be k _B.

Similarly to the case of the first embodiment, the transfer function of equation (9) is also represented by (1-μ _z z ^-1 ) / (1-μ _p z ^-1 ) of the first-order pole-zero type. Is the transfer function of. Where μ _z ,
μ _p is an independent filter coefficient having an absolute value smaller than 1, and in this case μ _z = η _B (k _B ), μ
_p = η _A (k _A ).

The conversion formula from the LPC coefficient to the PARCOR coefficient, which uses the Durbin method algorithm in reverse, is well known, and is described in detail in, for example, "Digital Speech Processing" (Tokai University Press, Furui).

Next, the post filter 1 in the present embodiment.
The flow of processing in 03 will be described using the flowchart shown in FIG. First, the transfer function F (z) = A (z) / of the spectrum envelope emphasis filter 112 including a pole-zero filter.
Parameters of the zero-side transfer function A (z) and the pole-side transfer function B (z) in B (z) are acquired (step S2).
1). Next, in the parameter calculation units 123 and 23, Durb
Using the inverse algorithm of the in method, the parameter k _A of the first-order filter b (z) is changed from the parameter of _A (z) to B
From the parameters of (z), the parameters k _B of the primary filter a (z) are obtained by calculation (step S2
2), these a (z) and b (z) are set as parameters of D (z) as shown in Expression (9) (step S2).
3). The adaptive filter 121 performs a filtering process according to the transfer function D (z), thereby performing a process of compensating for the slope of the spectrum in the spectrum envelope emphasis filter 112.

Incidentally, 1 described in the first and second embodiments
The specific configuration of the second-pole / zero-type adaptive filter 121 can be represented by, for example, the signal flow of FIGS. 4 and 5. According to this embodiment, transmitting seek mu _p from the function A (z), by adopting a configuration for determining the mu _z from the transfer function B (z), a lower filter coefficients of orders, i.e. less The slope of the spectrum can be compensated by the amount of calculation.

Next, a third embodiment will be described. First
In the second and second embodiments, the compensation filter 113 is configured by using the parameters obtained based on the first-order prediction of poles and zeros, mainly for the purpose of compensating the slope of the spectrum provided by the spectrum envelope emphasis filter 111. Described how.

The third embodiment will explain that it is possible to compensate not only the slope of the spectrum but also the unevenness of the spectrum by using a method based on higher-order prediction. The present embodiment is characterized in that the primary prediction in the first and second embodiments is a secondary or higher prediction, and the external appearance is basically a diagram showing the first and second embodiments. The same as 1. The effect of introducing high-order prediction as in the present embodiment will be described below.

If the compensation filter 113 is constructed by using second-order prediction coefficients for each pole and zero, it is possible to weaken a part of the characteristic of the spectrum envelope emphasizing filter 111 for emphasizing the unevenness of the spectrum envelope. The reason is based on the nature of the prediction filter. That is,
The part of the spectral envelope where the emphasis is weakened is the frequency band near the first formant, which is most strongly emphasized by the ordinary post filter. Therefore, when the compensating filter 113 is configured using the second-order prediction coefficient, there is an effect that the formants in other frequency bands that are difficult to be emphasized by the ordinary post filter can be emphasized preferentially. By further increasing the order of the prediction coefficient, it becomes possible to emphasize the unevenness of the spectrum envelope of the voice in a frequency in a narrower range than when the second-order prediction coefficient is used. By using this method, it becomes possible to relatively easily emphasize a high-frequency formant of a vowel, which is difficult to be emphasized by a conventional post filter, without using a bandpass filter.

In the present embodiment, not only the slope compensation of the spectrum envelope emphasis filter, but also a higher-order spectrum slope that compensates for unnecessary spectrum slope (called pitch slope) brought about by using the pitch component emphasis filter. The compensation method will be described. The pitch component emphasis filter may be used in the post filter as shown in FIG. 1 or may be used in the audio signal reproduction unit. Here, in the audio signal reproduction unit, the drive signal of the synthesis filter is used. An example of using a pitch component emphasis filter will be described.

The waveform a in FIG. 6 is a diagram showing the shape of the spectrum of the drive signal of the synthesis filter and its slope (represented by a straight line in the figure for simplicity) in the current voice section.
As shown in the waveform a, the spectrum of the drive signal having the pitch periodicity has a harmonic structure having a spectrum peak at a frequency that is an integral multiple of the frequency corresponding to the pitch period.
Ideally, the slope of the spectrum envelope of the drive signal of the synthesis filter is flat, but when the spectrum of the actual drive signal is observed, there are many sections that cannot be said to be flat. The reason for this is that the spectrum envelope is not correctly analyzed, the synthesis filter cannot completely express the spectrum envelope of the speech, and the deterioration of the filter characteristic due to the lack of the number of coding bits of the synthesis filter in the speech coder. Can be considered.

CELP (Code Excited Linear Predicti
On the other hand, in the analysis / synthesis type speech coding apparatus such as the on) system, such a shortage of the characteristics of the synthesis filter is compensated by the characteristics of the drive signal. In such a case, it is obvious that the spectrum of the drive signal, which is originally flat, has an inclination or some unevenness. Further, the slope of the spectrum of the drive signal differs for each voice section (for example, frame or subframe).

The basic function of the pitch component emphasizing filter according to the prior art can be explained using the waveforms a, b and c of FIG. The waveform b shows an example of the shape of the spectrum of the drive signal of the synthesis filter and its inclination in the voice section temporally separated by the pitch period. The processing of the pitch component emphasis filter is performed by adding a signal of the current voice section by multiplying a signal temporally separated by the pitch period by the pitch gain β, thereby making the harmonic structure of the pitch clearer as shown in the waveform c. Is the process of Pitch gain β
Is determined according to the correlation of the drive signals separated by the pitch period.

However, the slope of the spectrum of the drive signal of the waveform a (expressed as Q (z) in the z conversion region as shown in the figure) is distant from this by a pitch period. As a result of emphasizing the pitch component by using the drive signal of the waveform b having a different spectrum inclination, the spectrum inclination of the drive signal of the waveform c after the pitch component emphasis is Q
The quality has changed from (z) to Q '(z).
That is, in this example, Q (z) shows an upward slope but Q '(z) has a downward slope.
According to the experiments by the inventors, although the conventional pitch component enhancement processing has an effect of reducing the noise sensation, such a change in the slope of the spectrum of the drive signal adds a muffled feeling to the sound or partially phonologically It has been found that the clarity of is lost. In particular, under the condition of the tandem connection in which the audio signal reproduced by the audio encoding / decoding process is encoded / decoded again and reproduced, the above-mentioned muffled feeling of sound and the partial unclearness of the phoneme are amplified. As a result, it is likely to be perceived as extreme sound quality deterioration.

In order to solve this problem, the feature of the present embodiment is to introduce a process of compensating for the inclination (or deformation) of the spectrum accompanying the pitch component emphasis process.
This compensation is performed by the conventional method for enhancing the pitch component filtering. The inclination of the spectrum Q '(z) of the drive signal of the waveform c is obtained.
Is to recover only the inclination to the original inclination Q (z) while maintaining the pitch harmonic structure as shown by the waveform d.
By doing so, the problems of muffled sound and phoneme deterioration caused by the pitch component emphasis filtering can be significantly improved.

That is, in the present embodiment, in order to restore the slope (or outline) Q '(z) of the spectrum deformed like the waveform c to the original slope (or outline) Q (z) of the spectrum. , Before or after the pitch component emphasis filtering process, the effect of Q ′ (z) is removed and Q
A process of adding the characteristic of (z), or Q (z) / Q '
The filtering process of (z) is performed. In order to realize this processing, it is necessary to extract at least the characteristic of Q (z).

FIG. 7 is a block diagram of a speech decoding apparatus according to this embodiment having a function of compensating for the inclination (pitch inclination) of the spectrum of the drive signal due to such pitch component emphasis filtering. In this voice decoding device, the configurations of a voice signal reproducing unit 102 'and a post filter 103' are different from those in FIG. The audio signal reproducing unit 102 'is
Before the drive signal is input to the synthesis filter to synthesize the audio signal, the pitch component enhancement filter is used to enhance the pitch component of the drive signal. That is, in this embodiment, the pitch enhancement filter provided in the post filter 103 of FIG. 1 is incorporated in the audio signal reproducing unit 102 ′, and the post filter 103 ′ is provided in the post filter 103 of FIG. The pitch component emphasizing filter 111 that has been used is removed.

FIG. 8 shows the audio signal reproducing section 10 in FIG.
It is a block diagram which shows the detailed structure of 2 '. The audio signal reproducing unit 102 'includes a synthesis filter information generating unit 201,
The drive signal generation unit 202, the first synthesis filter 203, the pitch component enhancement filter 204, the pitch inclination compensation filter 205, the first and second LPC analysis units 206 and 37, and the second synthesis filter 208 are included. The synthesis filter information generation unit 201 and the driving signal generation unit 202 determine the filter coefficients of the synthesis filters 203 and 208 from the parameter information decoded by the parameter decoding unit 101 in FIG. 7, respectively, and the first synthesis filter. The drive signal e (n) of 203 is generated.

The drive signal e (n) generated by the drive signal generation unit 202 is input to the first synthesis filter 203, and the pitch component enhancement filter 204 and the first L filter are added.
It is input to the PC analysis unit 206. The drive signal ep in which the pitch component is emphasized by the pitch component emphasis filter 204.
(N) is input to the pitch inclination compensation filter 205 and the second LPC analysis unit 207. First and second LPC
In the analysis units 206 and 207, the pitch tilt compensation filter 2
A filter coefficient of 05 is generated. The pitch gradient compensation filter 205 compensates the pitch gradient, that is, the drive gradient of the spectrum of the drive signal compensated by the pitch component enhancement filter 204 is input to the second synthesis filter 208.
The audio signal is reproduced. The reproduced audio signal is further input to the spectral envelope emphasis filter 112 in the post filter 103 '. Further, the synthesis filter information generated by the synthesis filter information generation unit 201 is the transfer function F shown in the equation (1) of the spectrum envelope emphasis filter 112.
Used to determine (z). Further, the output signal of the first synthesis filter 203 is used for the gain determination of the gain adjusting unit 114 in the post filter 103 '.

Next, the pitch component emphasizing filter 204, the pitch inclination compensating filter 205 and the first,
The second LPC analysis units 206 and 207 will be described in more detail.

The first LPC analysis section 206 performs an L-th order linear prediction analysis on the drive signal e (n) of a predetermined section of the reproduced audio signal, for example, one subframe or one frame section, and predicts L predictions. Find the coefficient. Since the method of linear prediction analysis is well known, detailed description is omitted here. Prediction coefficient [rho ₁ in the case of L = 1 can be obtained by _{ρ 1 = Σe (n) e} (n + 1) / Σe (n) e (n) (12). When L = 1, the slope characteristic Q (z) of the spectrum described in FIG. 6 can be expressed as Q (z) = 1 / (1-g (ρ ₁ ) z ⁻¹ ) (13). Here, g () is a function for adjusting the prediction coefficient. As an example, g (ρ ₁ ) = ηρ
_{For 1} and η, values greater than 0 and less than 1 are used. When L is 2 or more, there is an effect that a more detailed outline of the spectrum of e (n) can be expressed by Q (z). At this time Q (z)
Can be expressed as Q (z) = 1 / (1-ρ ₁ z ⁻¹ −ρ ₂ z ⁻² −... −ρ _L z ^−L ). Where ρ ₁ , ρ ₂ , ..., ρ _L are L
It represents L prediction coefficients obtained from the following linear prediction analysis.

The pitch component emphasizing filter 204 receives the drive signal e (n) and pitch-enhances the drive signal ep.
(N) is output. As a method of this pitch emphasis filtering, for example, ep (n) = e (n) + βe (n−T), n = 0, 1, ..., N−1 (14) can be used. Here, T is the pitch period, N is the length of the section used for pitch enhancement, and β is the pitch gain. The value of β can be determined based on the value obtained by the pitch analysis, and a value of 0 <β <0.7 is usually used. Further, as another method, a method of using a fixed value prepared in advance for β depending on the presence or absence of the pitch period is also effective. -For example, β = 0 when there is no pitch period,
When the pitch periodicity is relatively strong, β = 0.6
Determine the value of.

The second LPC analysis section 207 performs M-th order linear prediction analysis on the pitch-enhanced drive signal ep (n) to obtain M prediction coefficients. The prediction coefficient ρ ₁ ′ when M = 1 can be obtained by ρ ₁ ′ = Σep (n) ep (n + 1) / Σep (n) ep (n) (15).

When M = 1, the slope characteristic Q '(z) of the spectrum explained in FIG. 6 is expressed as Q' (z) = 1 / (1-f (ρ ₁ ′) z ⁻¹ ) (16). be able to. Here, f () is a function for adjusting the prediction coefficient. -As an example, f (ρ ₁ ′) =
For η′ρ ₁ ′ and η ′, values larger than 0 and 1 or less are used. When M is 2 or more, there is an effect that a more detailed spectrum outline of ep (n) can be expressed by Q '(z). At this time, Q ′ (z) can be expressed as Q ′ (z) = 1 / (1-ρ ₁ ′ z ⁻¹ −ρ ₂ ′ z ⁻² −... ρ _M ′ z ^−M ) (17). Where ρ ₁ ', ρ ₂ ', ..., ρ
_M ′ can represent M prediction coefficients obtained from an M-th order linear prediction analysis.

The pitch inclination compensation filter 205 is an LPC.
Q '(z) based on the prediction coefficients from the analysis units 206 and 207.
And Q (z) are used to perform a filtering process to give the characteristic of Q (z) / Q '(z) to the input drive signal ep (n) after the pitch component emphasis, and a signal eq in which the pitch inclination is compensated is performed. (N) is given to the second synthesis filter 208. L
= 1 and M = 1, using equations (13) and (16), Q (z) / Q ′ (z) = (1-f (ρ ₁ ′) z ⁻¹ ) / (1-g ( ρ ₁ ) z ⁻¹ ) (18). When η and η ′ are used, η = η ′ = 1
Then, Q (z) / Q ′ (z) = (1-ρ ₁ ′ z ⁻¹ ) / (1-ρ ₁ z ⁻¹ ) (19) can be expressed.

FIG. 9 is a more detailed view of Q (z) and Q '(z) for the case where L = 1, M = 1, and η = η' = 1 in the principle diagram of the spectrum tilt correction shown in FIG. It is the figure shown in a typical form.

Returning to FIG. 8, the description of the audio signal reproducing unit 102 'will be continued. When the drive signal eq (n) after the pitch inclination compensation is given to the second synthesis filter 208, the one adjusted so that the power of eq (n) is about the same as the power of eq (n) is newly added. The method of giving it to the synthesis filter 208 as n) is also effective. The second synthesis filter 208 synthesizes the drive signal eq (n) in which the pitch inclination, that is, the spectrum inclination due to the emphasis of the pitch component is corrected, and reproduces the audio signal in which the pitch component is emphasized. The reproduced audio signal is given to the post filter 103 '. In order to give power information from the audio signal reproducing unit 102 'to the gain adjusting unit 114 of the post filter 103', here, the drive signal e (n) generated by the drive signal generating unit 202 is input to the first synthesis filter 33. Then, an audio signal in which pitch enhancement is not performed is obtained. When the drive signal eq (n) whose power has been adjusted as described above is used, the second synthesis filter 38 is used without using the first synthesis filter 203.
It is also effective to provide the gain adjusting section 411 with an audio signal in which the pitch component, which is the output of, is emphasized.

Next, the flow of processing in this embodiment is shown in FIG.
This will be described using the flowchart of 0. First, the drive signal generation unit 202 generates the drive signal e (n) of the first synthesis filter 203 (step S31), and the drive signal e (n) is subjected to the primary self-phase by the first LPC analysis unit 206. The relation number ρ ₁ is obtained (step S32). Also,
The drive signal e (n) is input to the pitch component emphasis filter 204 to obtain the drive signal ep (n) in which the pitch component is emphasized (step S33), and the second LPC analysis unit 207 for this drive signal ep (n). And the first-order autocorrelation coefficient ρ ′
₁ is obtained (step S34). Using these autocorrelation coefficients ρ ₁ and ρ ′ ₁ , the pitch tilt compensation filter 205 compensates the pitch tilt, that is, the tilt of the spectrum of the drive signal ep (n) in which the pitch component is emphasized (step S35). Next, the drive signal eq (n) in which the pitch inclination is compensated is input to the second synthesis filter 208 to perform synthesis filtering, thereby reproducing the audio signal. The above steps S31 to S35 are the processing of the audio signal reproducing unit 102 '.

Next, the audio signal reproducing section 1 is operated as described above.
The audio signal reproduced in 02 'is post-filtered 103'
In the same manner as in the previous embodiment, the spectrum envelope emphasizing filter 112 first performs the spectrum envelope emphasizing filtering (step S37), and the compensation filter 113 subsequently corrects the slope of the spectrum generated by the spectrum envelope emphasizing filtering (step S37).
38). Finally, the gain adjusting unit 114 smoothly adjusts the gain so that the audio signal after processing by the post filter 103 'has a power equivalent to that of the audio signal before processing, and outputs the final audio signal (step S39).

As another method of realizing the fourth embodiment, the slope (or approximate shape) Q (z) of the spectrum of the drive signal before pitch enhancement in the current section is extracted and the signal used for pitch enhancement is extracted. It is also possible to realize the processing by giving the characteristic of Q (z) to the drive signal after the pitch component emphasis by filtering the pitch component emphasis after flattening the slope of the spectrum included in the above. Further, as a method for more stably adjusting the pitch tilt compensation, Q (z / γ) instead of Q (z),
Q ′ (z / γ ′) may be used instead of Q ′ (z). Here, γ and γ ′ are, for example, 0 <γ <
It can be set to 1, 0 <γ ′ <1.

Next, a fifth embodiment will be described. The present embodiment is an example of performing the slope compensation processing of the spectrum by using the adaptive filter of the transfer function Tpz (z) which is a further improvement of the adaptive filter of the transfer function D (z) described in the second embodiment. This has the effect of improving the clarity of the consonant section and improving the crispness of the sound.

FIG. 11 shows an embodiment in which the post filter according to the present embodiment is applied to the final stage of a speech decoding apparatus, and blocks having the same functions as in FIG. 1 are assigned the same reference numerals. That is, the reproduced audio signal S is input from the encoded data (parameterized audio compression information) transmitted from the audio encoding device on the transmitting side input to the input terminal 100 via the parameter decoding unit 101 and the audio signal reproducing unit 102. (N) is reproduced, this reproduced audio signal is passed through the post filter 103, and the final output audio signal So is obtained.
(N) is generated. Hereinafter, the post filter 103 in this embodiment will be described in detail.

The post filter 103 includes a pitch component emphasis filter 111, a spectrum envelope emphasis filter 112,
The compensating filter 113 and the gain adjusting unit 114 are included, and each of these elements is configured as follows.

As described above, the spectral envelope enhancement filter 112 has a transfer function represented by the configuration of F (z) = A (z) / B (z), but here, the spectral envelope enhancement filter 112. In order to clarify the processing performed in the above, the processing will be divided into more detailed processing blocks.

10 input from the audio signal reproducing unit 102
LPC coefficients (here, a 10th-order LPC coefficient is used) are A (z) parameter calculation units 2200 and B.
(Z) is input to the parameter calculation unit 2201, and each of these parameter calculation units 2200 and 2201 has a parameter Awi (i = 1 to 10) of A (z) and a parameter bwi (i = 1 to 1) of B (z). 10) is calculated and output.

On the other hand, the signal input to the post filter 103 is subjected to the process of emphasizing the repetition of the pitch period by the pitch emphasizing filter 111, and then the zero type having the transfer function of A (z) in the spectrum envelope emphasizing characteristics. 1 / B after filtering by the filter 2202
It is filtered by the polar filter 2203 having a transfer function of (z).

Thus, the spectrum envelope emphasis filter 11
In the audio signal whose spectrum envelope is emphasized by 2, the unnecessary slope of the spectrum is further compensated by the compensation filter 113. In the compensation filter 113, the transfer function Tpz (z) of the adaptive filter 2121 that performs a specific filtering process is as follows: Tpz (z) = (1-μ _zero z ⁻¹ ) / (1-μ _pole z ⁻¹ ) (20 ) Is represented by. That is, this adaptive filter 2121 is also z
The transfer function of the conversion domain is (1-μ _z z ^-1 ) / (1-μ _p z ^-1 ) (where μ _z and μ _p are mutually independent filter coefficients with absolute values smaller than 1) The first-order pole-zero type shown is the same as in the previous embodiment.

In the filtering process by the adaptive filter 2121, first, the two filter coefficients μ _zero and μ _pole which determine the characteristics of the adaptive filter 2121 must be obtained in advance. These filter coefficients μ _zero ,
Each mu _pole is as follows mu _zero calculator 2124,
It is separately calculated by the μ _pole calculation unit 2123.

The μ _pole calculation unit 2123 inputs the parameters of A (z) which are the output of the parameter calculation unit 2200 of A (z), _{obtains an} autocorrelation coefficient r _1zero described later from these parameters, and μ _pole ′ = C ₀ r _1zero (r _1zero <Th) μ _pole ′ = C ₁ r _1zero (r _1zero ≧ Th) (21) μ _pole = C ₂ μ _pole ′ + (1-C ₂ ) last_μ _pole (22) Calculate μpole.

Here, the weighting factors C ₀ , C ₁ , C ₂ and the threshold value Th are adjustment values, and 0 <C ₁ <C ₀ ≦ 1, 0
<C ₂ ≦ 1, and Th is a value close to 0. last_μ
_pole is the μ of the immediately preceding speech section (for example, the previous subframe)
Represents a _pole . r _1zero is a spectral envelope enhancement filter 1
Among them, the first-order autocorrelation coefficient (equal to the first-order PARCOR coefficient) calculated by using the filter coefficients aw1 to aw10 of the zero-type filter 2202 having the transfer function A (z) on the numerator side. The value of r _1zero can be obtained as an autocorrelation value when one sample time is shifted using an impulse response sequence of 1 / A (z), but as a more efficient method, the above-mentioned Durbin recursive algorithm is used. (Or Levinson's recursive algorithm,
Alternatively, by using the Levinson-Durbin algorithm) in reverse, the primary autocorrelation coefficient can be obtained with a small amount of calculation without actually calculating the impulse response.

On the other hand, the μ _zero calculation unit 2124 uses B (z)
The parameter of B (z) which is the output of the parameter calculation unit 2201 of 1 is input, and the autocorrelation coefficient r _1pole is _obtained from them. The coefficient μ _zero is calculated by μ _zero = C ₃ r _1pole (23).

Here, C ₃ is the adjustment value of the weighting coefficient,
It is desirable that 0 <C ₃ <1. r _1pole is a first-order autocorrelation coefficient (first-order PARC) calculated using the coefficients bw1 to bw10 of the polar filter having the transfer function B (z) on the denominator side of the spectral envelope emphasis filter 112.
Is equal to the OR coefficient). The value of r _1pole is 1 / B
It can be obtained as an autocorrelation value when one sample time is shifted using the impulse response sequence of (z), but as a more efficient method, the above-mentioned Durbin's recursive algorithm (or Levinson's recursive algorithm, or Levinson's) is used. -Durbin algorithm) in reverse, the r _1pole can be calculated with a small amount of calculation without actually calculating the impulse response.
The value of can be obtained.

In the experiments conducted by the inventors, the above adjustment values were set to C
₀ = 0.9, C ₁ = 0.4, C ₂ = 0.7, Th = 0.
It was found that the sound quality was significantly improved when the values of 0 and C ₃ = 0.7 were set. Substituting these values ((2
1), (22) and (23) are written, μ _pole ′ = 0.9r _1zero (r _1zero <0.0) μ _pole ′ = 0.4r _1zero (r _1zero ≧ 0.0) ( a _{24) μ pole = 0.7μ pole '} + 0.3last_μ pole (25) μ zero = 0.7r 1pole (26).

The adaptive filter 2121 constitutes an adaptive filter of a pole first-order / zero-first-order transfer function Tpz (z) using the coefficients calculated as described above, and outputs a speech signal in which the spectrum envelope is emphasized. Filter as input.

Finally, the gain adjusting unit 114 smoothly adjusts the gain so that the output audio signal after being processed by the post filter 103 has the same power as the input audio signal before being processed, and the output of the post filter 103 is output. Audio signal.

Next, the post filter 1 in the present embodiment.
The flow of processing in 03 will be described using the flowchart in FIG. First, the spectrum envelope emphasis filter F (z)
Filter A (z) that constitutes (= A (z) / B (z))
Parameters awi (i = 1 to 10) of B and parameters bwi (i = 1 to 10) of B (z) are acquired (step S).
51). An example of a specific method of step S51 is an LPC coefficient αi in the current voice section from the voice signal reproducing unit 102.
Using (i = 1 to 10) awi = (γ1) ⁱ α _i (i = 1 to 10) (27) bwi = (γ2) ⁱ α _i (i = 1 to 10) (28) Is. At this time, A (z) and B
(Z) can be expressed as A (z) = 1 + Σawiz ⁻ⁱ (i = 1 to 10) (29) B (z) = 1 + Σbwiz ⁻ⁱ (i = 1 to 10) (30). When the positive and negative definitions of the LPC coefficient are different, equations (29) and (30) are: A (z) = 1−Σawiz ⁻ⁱ (i = 1 to 10) (29 ′) B (z) = 1 -Σbwiz ⁻ⁱ (i = 1 to 10) (30 ′). γ1 and γ2 are parameters for adjusting the degree of emphasis of the spectrum, and values having a relation of 0 <γ1 <γ2 <1 are usually used.

Next, filtering for pitch enhancement of the input voice signal (step S52) and filtering for spectral envelope enhancement (step S5).
Perform 3).

Next, Tpz (z) which characterizes this embodiment
Using the adaptive filter of the transfer function, the inclination of the spectrum is corrected as follows. First, the autocorrelation coefficient r1 _zero is obtained from the parameter awi (i = 1 to 10) of A (z) (step S54), and then r1 _zero and the threshold Th.
Is compared (step S55), and r1 _zero is T
If it is smaller than h, the _value multiplied by C ₀ is set as μ _pole ′ (step S56), and if Th or more, the value obtained by multiplying r1zero by C ₁ is set as μ _pole ′ (step S57). A value obtained by interpolating μ _pole ′ and last_μ _pole corresponding to the previous μ _pole using C ₂ is set as the μ _{pole of the} current speech section (step S5).
8). The obtained μ _pole value is stored in last_μ _pole for interpolation processing in the next voice section (step S59).

Next, the parameter bwi (i =
1 to 10), the autocorrelation coefficient r1 _pole is obtained (step S60), and the r1 _pole is multiplied by C ₃ to obtain μ _zero (step S61).

The two filter coefficients μ thus obtained
By performing the filtering process with the adaptive filter of the transfer function Tpz (z) determined by _pole and μ _zero , the unnecessary spectrum inclination mixed by the spectrum envelope emphasis filtering is compensated (step S62).

Finally, the gain is adjusted so that the output audio signal after processing by the post filter 103 has the same level of power as the input audio signal before processing, and the output is output from the post filter (step). S6
3).

Note that the adaptive filter used in this embodiment may implement the processing in such a way that it has its own filter gain. In that case, the transfer function Tpz (z) of the adaptive filter can be expressed as Tpz (z) = Gpz (1-μ _zero z ⁻¹ ) / (1-μ _pole z ⁻¹ ) (31). As the filter gain Gpz, Gpz = (1-λ _pole μ _pole ) / (1-λ _zero μ _zero ) (32) can be used. Here, λ _pole and λ _zero are fixed adjustment values that satisfy 0 <λ _pole and λ _zero <1.

By doing so, the adaptive filter of the transfer function Tpz (z) can be made to have a simple gain self-adjusting function, so that a compensating filter for compensating the slope of the spectrum is arranged after the gain adjusting section. This is effective in the case of such a post filter configuration.

As described above, according to the present embodiment, in addition to the effect of the previous embodiment, the first coefficient obtained from the parameter of A (z) using the weighting coefficient of the relation of C ₁ <C ₃ <C ₀ . When the autocorrelation coefficient r1 _zero of r1 _zero is smaller than a threshold value close to 0 (Th), r1 _zero is weighted with a weighting coefficient C ₀ ,
When 1 _zero is larger than the threshold Th, the value obtained by weighting r1 _zero with the weighting coefficient C ₁ is μ _pole
Is obtained and μ _zero is obtained from a value obtained by weighting the second autocorrelation coefficient r1 _pole obtained from the parameter of B (z) with the weighting coefficient C ₃ , whereby r1 _zero becomes smaller than the threshold Th. Is small, the voice in this section is a voice like a strong high-frequency consonant, and conversely r1 _zero is a threshold T.
If it is larger than h, the speech in this section is a speech like a strong vowel in the low frequency range. Therefore, by comparing the autocorrelation coefficient with the threshold Th as described above, the weighting coefficient is switched, and the compensation filter It is possible to make 113 the compensation characteristic suitable for each consonant and vowel, and to improve the sound quality more effectively.

Next, as a sixth embodiment, a post filter with an improved gain adjusting section will be described. FIG.
Shows an embodiment in which the post filter according to the present embodiment is applied to the final stage of a speech decoding apparatus, and blocks having the same functions as in FIG. 1 are assigned the same reference numerals. That is, the reproduced audio signal S is input from the encoded data (parameterized audio compression information) transmitted from the audio encoding device on the transmitting side input to the input terminal 100 via the parameter decoding unit 101 and the audio signal reproducing unit 102. (N) is reproduced and passed through the post filter 403 to generate the final output audio signal So (n). Hereinafter, the post filter 403 in this embodiment will be described in detail.

The post filter 403 comprises a filter processing section 410 and a gain adjusting section 414. The filter processing unit 410 performs various filtering processes in the post filter 403. Specifically, the spectrum envelope emphasis filter processing, the pitch emphasis filter processing, and the spectrum used together with the spectrum envelope emphasis filter processing performed based on the information such as the LPC coefficient α _i (i = 1 to 10) and the pitch period from the audio signal reproducing unit 102. Is a slope compensation filter process of. However, the filter processing unit 410 may not include all of the processes described here, and may not include the pitch enhancement filter process, for example.

The filter processing section 410 finds the zero-input response Z _i (n) and the zero-state response Zs (n) of the filter by the length corresponding to the current speech section, and these are obtained.
It outputs to 14. The zero input response Zi (n) is a response that is output depending on only the internal state of the filter when the filter is operated assuming that the signal on the input side of the filter processing unit 410 is completely zero. On the other hand, the zero-state response Zs
(N) is a response output when an input is given to the filter processing unit 410 assuming that the internal state of the filter is zero.

The gain adjusting unit 414 is the gain calculating unit 41.
5, the gain multiplication unit 416 and the addition unit 417, and the zero-state response Zs (n) from the filter processing unit 410.
The gain calculation unit 415 calculates the gain to be multiplied by, and the gain multiplication unit 416 multiplies the gain, and the addition unit 417 adds the result to the zero input response. As a result, the power-adjusted output audio signal So (n) is generated and output to the audio signal output terminal 404.

When the gain adjusting method according to the present embodiment is used, the power of the output audio signal So (n) of the post filter 403 is completely combined with the power of the input audio signal S (n) in units of a predetermined audio section (for example, subframe). Can be matched to. Moreover, it is possible to avoid discontinuity in the power of the output audio signal at the boundary of the section without performing any processing such as gain smoothing. Further, in the present embodiment, it is determined whether or not the power can be matched when a positive gain is used, and if not, the gain is set to a gain value C ₄ (≧ 0) which is less affected by the power mismatch in input and output. To be done. This has the effect of stably improving the sound quality of the output audio signal So (n) from the post filter 403.

The gain calculator 415 calculates the gain g based on the following equation. IF (d> 0) (33) g = [sqrt (b ² + d) -b] / a (34) else g = C ₄ (35) endif where a = ΣZ _s (n) Z _s (n) (N = 0 to N−1) (36) b = ΣZ _i (n) Z _s (n) (n = 0 to N−1) (37) d = a (ΣS (n) S (n) −ΣZ a _{i (n) Z i (n} )) (n = 0~N-1) (38). Where the function sqrt (x) is the square root of x, N
Represents the length of a predetermined voice section (for example, a subframe). The parameter C ₄ is a value used as g in a bad state where the powers of the input and output audio signals cannot be matched with a non-negative gain, and is a numerical value in the range of 0 ≦ C ₄ <1. desirable. For example, C ₄ = 0.5
You may make it a fixed value like this.

Based on the condition (d> 0) of the equation (33), g
When g is obtained, it is possible to reliably prevent g from becoming a negative value, and there is an advantage that stable gain adjustment can be realized. This condition means that the power of the zero-state response is positive and the power of the input voice signal is larger than the power of the zero-input response, as can be seen from the equations (36) and (38). If this condition is not satisfied, the input and output powers cannot be matched with a positive (plus) gain.

The above (34), (36), (37),
Expression (38) is also shown in Japanese Patent Application No. 2-41286 (adaptive post filter), but this method has a problem in the conditional expression for obtaining the gain g. That is, Japanese Patent Application No. 2-412
In 86, "If the value b ² + d in the brackets of sqrt is positive, g is calculated from the equation (34)".
The value of g obtained from this may be negative. If a negative gain value is used, the waveform after the zero-state response Zs (n) is multiplied by the gain will be inverted, and the shape of the final output speech waveform will be disturbed, resulting in a jarring noise that is apt to sound in terms of sound quality. There was a problem that was mixed.

This will be explained using a specific numerical example. From equations (35), (36) and (37), a = 2 and b =
5 and d = −24, Japanese Patent Application No. 2-4128
In the method described in No. 6, b ² + d = 5 ² -24> 0.
Therefore, using the gain calculation formula (34), g =
Becomes ^{(sqrt (5 2 -24) -5} ) / 2 = -2 , and the called forced match input and output power by deforming the waveform using a negative gain is achieved results.

On the other hand, in this embodiment, since d is a negative value, the conditional expression of the equation (33) does not use the equation (34), but the positive gain value g = C ₄ (1 > C ₄ ≧
0). As described above, in the gain adjustment in the present embodiment, if the input and output powers are not matched with a negative gain and the powers cannot be matched with a positive gain, the influence of the gain mismatch is reduced as much as possible. As described above, the gain g is replaced with a non-negative value C ₄ . As a result, there is an effect that the sound quality of the post filter can be improved more stably than before.

FIG. 14 shows an example of a signal flow of a more detailed process of the gain calculation section 415. In the figure, the calculation unit 420 calculates the power from the input audio signal S (n) (corresponding to the first term in parentheses on the right side of the expression (38)). The calculation unit 421 calculates the power of the quiescent response Z _i (n) (corresponding to the second term in the parentheses on the right side of Expression (38)). The calculator 422 calculates the power of the zero-state response Z _s (n) (corresponding to a in the equation (36)). The calculator 423 calculates the inner product of the zero-input response and the zero-state response (corresponding to b in equation (37)). Next, the gain determination unit 425 determines the calculation units 420 and 42.
Based on the calculated values (information of parameters a and d) from 1, 422, the condition corresponding to the equation (33) is determined. However,
The parameter b of the equation (37) is not used for the determination. Based on this determination result, the gain determining unit 42 is provided with the determination information whether to use the equation (34) or the equation (35) for the gain calculation.
Pass to 6. The gain determination unit 426 includes calculation units 420 and 42.
The calculated values from 1, 422 and 423 and the positive gain value C ₄ from the positive gain output unit 424 are input, and the gain is calculated from the formula (34) or (35) according to the judgment information from the gain judgment unit 425. Is determined as g and is used as the output of the gain calculation unit 415.

Returning to FIG. 13 and continuing the description, the gain multiplication unit 416 multiplies the zero-state response Z _s (n) input from the filter processing unit 410 by the gain g calculated by the gain calculation unit 415. The addition unit 417 is a multiplication unit 416.
Output signal and quiescent response Z from the filter processing unit 410
_The signal obtained by adding _i (n) is output to the output terminal 404 of the post filter as the output audio signal So (n). The output of the gain adjusting unit 414, that is, the output So (n) of the post filter 403 can be expressed by the following equation.

So (n) = Z _i (n) + gZ _s (n) (n = 0 to N−1) (39) In this embodiment, unlike Japanese Patent Application No. 2-41286, (39)
The gain g of the expression is always a value of zero or more. Therefore, Z _s
Since it is possible to reliably prevent the waveform of (n) from being inverted,
It is possible to provide a post filter in which the sound quality of the output audio signal So (n) is stably improved.

Output audio signal So obtained by the equation (39)
Of (n), the latter half P (So (NP), ..., So)
Since the value of (N-1)) can be used as the initial internal state of the filter used for the zero-input response calculation in the next speech section, the filter processing unit 410 is notified of this So (n) as shown in FIG. Data 418 indicating the latter half P values is passed.

Next, the flow of processing performed in one voice section in this embodiment will be described with reference to the flowchart of FIG. First, the parameterized voice compression information is decoded (step S71), and the voice signal S (n) is reproduced from this decoded information (step S72). The audio signal S (n) is input to the post filter, and the LPC coefficient and pitch information necessary to configure the filter in the post filter are input to the post filter (step S73). Next, the processing in the post filter is started. First, a zero-input response and a zero-state response are obtained by the filter processing unit inside the post filter 403 (step S74).
Next, using the zero input response, the zero state response, and the input voice signal, the parameters a and d necessary for the gain determination are (3
It is calculated by the equations (6) and (38) (step S7).
6). The parameter d of the calculated parameters a and d is subjected to the gain determination of the equation (33) (step S7).
7) If the condition is met (Yes), then (3
The gain g is obtained by using the equations (7) and (34) (steps S78 and S79). If the condition is not satisfied (No), the gain is g = C4 by using the equation (35).
Is set (step S80). An output audio signal So (n) is obtained by adding the signal obtained by multiplying the zero-state response by g and the zero-input response (step S81). Finally, So (n) is used to update the initial internal state of the filter used for the zero-input response calculation (step S82).

As described above, according to this embodiment, in order to compensate the power change of the audio signal due to the filtering process performed on the audio signal to adjust the spectrum shape of the audio signal, the gain by which the audio signal is multiplied is set. When adjusting, the gain to be multiplied to the audio signal is calculated, and then the positive / negative judgment of the gain is performed. If this is negative, the gain is given to a non-negative value given by a predetermined method, preferably a small value of 0 or more and less than 1 By replacing, it is possible to avoid the sound quality deterioration due to the negative gain.

In this embodiment, the gain adjustment is performed by adjusting the power of the output audio signal So (n) using the power of the input audio signal S (n) of the gain adjusting unit as an index as shown in the equation (38). However, the index used for the gain adjustment is not limited to the power of the input audio signal. For example, the gain is made different depending on the power information obtained from the audio signal reproducing unit 102 and the voiced section and the silent section of the voice. The present invention is effective when the information for the above is used as an index for gain adjustment.

In the embodiment described above, F (z)
= (A (z) / B (z)) As a method of compensating for the unnecessary slope of the spectrum produced by the spectrum envelope emphasizing filter 112, (1) the spectrum slope produced by the transfer function Α (z) on the numerator side is The zero filter is used to compensate for the spectral slope caused by the transfer function B (z) on the denominator side by the polar filter (zero pole method), and (2) The spectrum of the numerator side transfer function Α (z) A method (pole-zero method) of compensating the gradient with a pole filter and compensating the gradient of the spectrum caused by the transfer function B (z) on the denominator side with a zero filter
Two methods have been described. As a method combining these methods (1) and (2), (3) the slope of the spectrum caused by the transfer function Α (z) on the numerator side and the transfer function B (z) on the denominator side respectively. There are possible methods such as a method of compensating with an adaptive filter that combines zero filters and zero filters (zero-zero method), and a method of compensating (4) using a combination of pole filters and pole filters (pole-pole method). The description is omitted.

In the above embodiment, the adaptive filter 12 is used.
1 and the filter coefficient of the pitch slope compensation filter 205 are
They are updated together with the filter coefficients of the spectrum envelope emphasis filter 112 and the pitch component emphasis filter 204. On the other hand, in order to update the filter coefficient more smoothly in time, the filter coefficient used in the previous speech section in the adaptive filter 121 and the pitch gradient compensation filter 205 and the spectrum envelope emphasis filter in the current speech section. 112 and the filter coefficient interpolated using the filter coefficient obtained from the filter coefficient of the pitch component enhancement filter 204 in the current speech section.
1 and the method used for the pitch slope compensation filter 205 is also effective. In this way, changes in the transfer function of the adaptive filter 121 and the pitch slope compensation filter 205 gradually change, so that it is possible to prevent a phenomenon in which the final audio signal slightly changes due to background noise.

Next, a seventh embodiment will be described with reference to FIGS. Although the first to sixth embodiments have described the examples in which the present invention is applied to the post filter on the decoding side, in the present embodiment, the weight filter used in the encoding side is the spectral shape adjusting method of the present invention. An example is shown used to compensate for the unwanted spectral tilt of the.

According to the present embodiment, by using the weight filter in which the slope of the spectrum is compensated, the weighting of the distortion measure, which is an index of code selection, is optimized, so that the code that can more faithfully express the original sound is selected. As a result, there is an effect that the sound quality can be improved even with the same bit rate and the same encoding method.

FIG. 16 shows an example in which the weight filter according to this embodiment is applied to a speech coder. The audio signal input to the input terminal 70 is analyzed and coded in frame units,
The finally encoded information is output from terminals 84-87. The encoding performed here is roughly performed to encode two pieces of information, that is, information of the synthesis filter and information of the drive signal. The information of the synthesis filter is often extracted from the voice and encoded for each frame unit having a length of about 10 to 30 ms, and the information of the drive signal is often encoded in subframe units obtained by dividing a frame into smaller parts. Here, for the sake of simplicity, an example will be described in which the drive signal is encoded in frame units.

As described above, the output signal obtained from the synthesis filter to which the drive signal is input becomes the reproduced audio signal. The speech coding apparatus of FIG. 16 will be described in detail below.

The speech coding apparatus shown in FIG. 16 includes a synthesis filter information analysis unit 71, a weight filter information calculation unit 72, a weight filter 73, a target signal generation unit 74, and an adaptive codebook 7.
5, noise codebook 76, gain codebook 77,
The gain adding units 78 and 79, the addition unit 80, the weighted synthesis filter 81, the distortion evaluation unit 82, and the code selection unit 83 are included. The weighting filter information calculation unit 72 further includes a WA calculation unit 88, a WB calculation unit 89, a μ _p calculation unit 90, and a μ _z calculation unit 91. In this embodiment, the μ _p calculator 90, μ
_The fact that the characteristic of the weighting filter is compensated based on the information obtained by the _z calculation unit 91 is different from the conventional speech coding.

The operation of the seventh embodiment will be described below with reference to the flowchart of FIG. The synthesis filter information analysis unit 71 analyzes the voice signal input from the terminal 70 on a frame-by-frame basis, and extracts the parameters of the synthesis filter capable of expressing the shape of the spectral envelope of the corresponding voice signal. LP is used as the method for extracting the parameters of the synthesis filter.
A method of extracting the LPC coefficient using C analysis can be used. The extracted parameters are converted into parameters suitable for quantization in the synthesis filter information analysis unit 71 and encoded. Information of the encoded synthesis filter is output to the terminal 84.

The information on the synthesizing filter quantized by the synthesizing filter information analyzing unit 71 is the weighted synthesizing filter 8
Used in 1. The information of the synthesis filter that is not quantized is used by the weight filter information calculation unit 72. The weight filter information calculation unit 72 calculates the parameters of the weight filter used in the weight filter 73 and the weighted synthesis filter 81 based on the information of the unquantized synthesis filter.
Here, an example will be described in which the information of the synthesis filter that is not quantized is used in this calculation, but a method of using the information of the quantized synthesis filter can also be used depending on the application. The weight filter W (z) used in this embodiment is represented by the following equation.

[0126]

[Equation 1] Here, WA (z) / WB (z) is a weighting filter of a conventional configuration, but since this includes an unnecessary spectrum slope, the pole-zero filter ( ^{_{1-μ z z -1) /}} (1-μ p z -1) is used to take a gist of the present embodiment. Here, a first-order pole-zero filter will be used for explanation, but the present invention is not limited to this. Further, in order to reduce the amount of calculation required for filtering the weight filter, a filter having a different configuration so that the characteristic of W (z) in the expression (40) can be maintained to some extent is used as W (z). The method is also effective and can be realized by the present invention. As a specific example, a time window is applied to the impulse response on the right side of the equation (40) to terminate the calculation with a predetermined short K + 1 sample. By using such a filter as the weighting filter W (z), it is possible to configure a weighting filter having a small calculation amount including the slope compensation processing of the spectrum shape of the present invention. At this time, the weight filter W (z) including the tilt compensation is

[0127]

[Equation 2] Here, window (i) represents a time window, and w (i) represents the impulse response of the transfer function on the right side of Expression (40). A square window or a Hamming window can be used as the time window.

The WA calculation unit 88 and the WB calculation unit 89 of the weight filter information calculation unit 72 obtain the parameters of WA (z) and WB (z), respectively, which form the weight filter.
An example of a specific calculation method is shown below. Assuming that the order of the LPC analysis is P, the LPC coefficient α _i (i
= 1 to P), the coefficient φ _i of WA (z) and WB
Ψ _{i of} (z) is calculated by φ _i = (υ ₁ ) ⁱ α _i (i = 1 to P) (42) ψ _i = (υ ₂ ) ⁱ α _i (i = 1 to P) (43) . A value of about 10 is often used for P in speech coding. here

[0129]

[Equation 3] Is. Υ ₁ and υ ₂ used in equations (42) and (43)
Is a parameter for adjusting the degree of weighting, and 0 <υ ₂ <
There is a relation of υ ₁ <1 (note that the adjustment value is different from that of the post filter). Typical values are υ ₁ = 0.9, υ
₂ = 0.4. The μ _p calculation unit 90 uses the information of WA (z) from the WA calculation unit (here, (φ _i ) to calculate WA
The coefficient μ _{p of the} polar filter for compensating for the unnecessary spectral tilt of (z) is obtained. That is, similar to the method described in the second embodiment, by using the inverse algorithm of the Durbin method, the first-order PARCOR from φ _i is used.
Find the coefficient and call it μ _p .

The μ _z calculation unit 91 uses the WB from the WB calculation unit.
Using the information of (z) (here, ψ _i ), the zero filter coefficient μ _z for compensating for the unnecessary spectrum inclination of WB (z) is obtained. That is, similar to the method described in the second embodiment, the inverse algorithm of the Durbin method is used to obtain the first-order PARCOR coefficient from ψ _i and set it as μ _z . In addition to this, a method of further modifying the values of the coefficients μ _p and μ _z used for the weight filter to perform more detailed optimization adjustment is also effective. As an example, the following coefficient transformation is performed.

Μ _p ← y _p μ _p (46) μ _z ← y _z μ _z (47) Here, y _p and y _z are coefficients for adjustment, and 0 <y _p <=
It is desirable that 1 and 0 < _yz <= 1. As a further optimization method, it is one of the optimization methods that the adjustment method of the pole / zero filter for compensation is changed according to the characteristics of WA (z) or WB (z) or the composite filter. For example, focusing on the characteristics of any one of the above filters, a method of adapting the value of the adjustment coefficient depending on whether this is a high-pass type or a low-pass type is also effective. Conceivable.

Referring again to FIG. 16, the information obtained by the weight filter information calculation unit 72 is used by the weight filter 73 and the weighted synthesis filter 81. The weight filter 73 weights the input voice signal based on the information obtained by the weight filter information calculation unit 72, and outputs the weighted voice signal to the target signal generation unit 74. Based on this weighted audio signal, the target signal generation unit 74 removes the influence of the encoding of the previous frame by the level of the weighted audio signal, and generates the target signal used for encoding the drive signal of the current frame.

Next, the drive signal is encoded. Three codebooks, an adaptive codebook 75, a noise codebook 76, and a gain codebook 77, are used to encode the drive signal. The adaptive codebook stores past drive signals, and represents a pitch period component of the drive signal by a pitch vector having a pitch period corresponding to the pitch period as a code.
The noise codebook represents the noise component of the drive signal by the noise vector corresponding to the noise code that is the code. Further, the gain code book is used for the purpose of controlling the magnitudes of the pitch vector and the noise vector, and the gain candidates corresponding to the gain code are given by the gain giving unit 78 and the gain giving unit 79, and the pitch after the gain is given is given. The addition unit 80 adds the vector and the noise vector to generate drive signal candidates. The drive signal candidates thus generated are passed through the weighted synthesis filter 81, and then the distortion evaluation unit 82 for calculating the distortion evaluation value corresponding to the distortion with the target signal.
Entered in. The code selection unit 83 searches each codebook for a code such that the calculated distortion evaluation value is small.

The above is the principle of the code search of the drive signal, but in order to reduce the calculation amount normally required for the search, there is a method of searching the code stepwise in the order of the adaptive codebook, the noise codebook, and the gain codebook. Mostly used.
Of the codes representing the drive signals finally searched, the pitch period (the code of the adaptive codebook), the noise code (the code of the noise codebook), and the gain code (the code of the gain codebook) are terminals 85 and 86, respectively. It is output to 87.

Next, the processing of the speech coding apparatus according to this embodiment will be described with reference to the flowchart of FIG. Initially, initial setting is performed (step S180). Next, the necessary amount of audio signals so that the processing for each frame can be performed is input to the synthesis filter information analysis unit 71 (step S1).
81). The synthesis filter information analysis unit 71 analyzes this voice signal, extracts the parameter of the synthesis filter corresponding to the voice signal, and encodes this (step S182). Next, weight filter information for forming the weight filter is generated (step S183). More specifically, this step calculates the coefficient of WA (z) (step S18).
4), calculation of μ _p using the coefficient of WA (z) (step S185), calculation of coefficient of WB (z) (step S18)
6), calculation of μ _z using the coefficient of WB (z) (step S187).

Next, a weighted voice signal is generated using the information of the obtained weight filter (step S18).
8) The influence of the encoding of the previous frame is removed by the weighted level of the audio signal, and the target signal used for encoding the drive signal of the current frame is generated (step S18).
9). The drive signal is encoded by sequentially performing adaptive codebook search (S191), noise codebook search (S191), and gain codebook search (S192) using this target signal. As the weight filter of the weighted synthesis filter used at this time, the one obtained in step S183 is used. Finally, the coded data of the current frame obtained by the coding is output.

In the above, μ _p is calculated from WA (z) and W
Although the example of obtaining μ _z from B (z) has been described,
From WB (z) by the method as shown in the embodiment of
It goes without saying that an embodiment in which _p is obtained and μ _z is obtained from WA (z) is also effective. It is also possible to use a pole-zero filter of the second or higher order as shown in the third embodiment.

Further, in the above embodiment, the order of various filters such as the pitch component emphasis filter, the spectrum envelope emphasis filter, the adaptive filter, and the pitch slope compensation filter can be arbitrarily changed, that is, these respective filters are connected in cascade. If you have. Further, the order of various filters used in the weighting filter is arbitrary.

Further, in the above embodiments, the example in which the present invention is applied to the final stage of the speech decoding apparatus has been described. However, in order to improve the subjective quality, other than the decoded speech signal in the speech encoding / decoding system. Also, the present invention can be applied to various voice signals such as a synthesized voice signal obtained by a voice synthesizer.

[0140]

As described above, according to the present invention, the first filter having the pole-zero type transfer function of the audio signal represented by A (z) / B (z) and the first filter By passing it through a second filter to compensate the characteristics of the filter,
When adjusting the spectral shape of the speech signal, by calculating the two parameters of the second filter separately from A (z) and B (z), it is possible to reduce the quality of the speech signal such as decoded speech or synthesized speech. The amount can be effectively improved.

According to the present invention, the second filter performs the filtering process with the polar filter and the zero filter having different parameters, so that the parameter is configured by the conventional first-order zero filter. Since the number of filters is increased as compared with the number of filters, the degree of freedom in expressing the transfer function of the filter is increased, the slope compensation of the spectrum can be flexibly performed, and the effect of improving sound quality is further enhanced. In this case, we seek mu _p from A (z), if the B (z) to determine the mu _z, it is possible to compensate for the tilt of the spectrum at lower filter coefficients of orders.

Further, using the weighting factors of the relationship of C ₁ <C ₃ <C ₀ , the first autocorrelation coefficient obtained from the parameter of A (z) is calculated from a threshold value (Th) which is close to 0. When it is small, the first autocorrelation coefficient is weighted with the weighting coefficient C ₀ , and when the first autocorrelation coefficient is larger than the threshold Th, the first autocorrelation coefficient is weighted with the weighting coefficient C ₁ . It is possible to obtain μ _p from the value obtained by performing the above, and to obtain μ _z from the value obtained by weighting the second autocorrelation coefficient obtained from the parameter of B (z) with the weighting coefficient C _3. For example, when the first autocorrelation coefficient is smaller than the threshold Th, the voice in this section is a strong consonant voice in the high frequency range, and conversely, the first autocorrelation coefficient has the threshold Th. If it is larger than the above, the sound in this section is a strong vowel sound in the low frequency range. Autocorrelation coefficient and threshold T
By switching the weighting coefficient by comparing with h, the second filter can have the compensation characteristics adapted to the consonant and the vowel respectively, and the sound quality is effectively improved.

Further, according to the present invention, in adjusting the gain for compensating the power change of the audio signal due to the filter processing for adjusting the spectrum shape of the audio signal, the positive / negative judgment of the gain to be multiplied by the audio signal is performed. When the gain is negative, by replacing the gain with a non-negative value given by a predetermined method, in particular, a small value such as 0 or more and less than 1, it is possible to avoid the sound quality deterioration due to the gain adjustment caused by using the negative gain. it can.

[Brief description of drawings]

FIG. 1 is a block diagram of a speech decoding apparatus incorporating post filters according to first to third embodiments.

FIG. 2 is a flowchart showing a processing flow of a post filter according to the first embodiment.

FIG. 3 is a flowchart showing a processing flow of a post filter according to the second embodiment.

FIG. 4 is a block diagram of an adaptive filter used in the present invention.

FIG. 5 is a block diagram of another adaptive filter used in the present invention.

FIG. 6 is a diagram for explaining the basic function of a pitch component emphasis filter and the principle of spectrum tilt compensation by pitch component emphasis processing.

FIG. 7 is a block diagram of a speech decoding apparatus incorporating a post filter according to a fourth embodiment.

8 is a block diagram of an audio signal reproducing unit in FIG.

9A and 9B are views for explaining the function of a pitch component emphasis filter and the effect of spectrum tilt compensation by the pitch component emphasis process in the fourth embodiment.

FIG. 10 is a flowchart showing the flow of processing in the fourth embodiment.

FIG. 11 is a block diagram of a speech decoding apparatus incorporating a post filter according to a fifth embodiment.

FIG. 12 is a flowchart showing the flow of processing in a post filter according to the fifth embodiment.

FIG. 13 is a block diagram of a speech decoding apparatus incorporating a post filter according to a sixth embodiment.

FIG. 14 is a block diagram showing the configuration of a gain calculation unit in FIG.

FIG. 15 is a flowchart showing a processing flow of a post filter according to the sixth embodiment.

FIG. 16 is a seventh embodiment and is a spectrum shape adjusting device for a voice signal applied to an encoding device.

17 is a flowchart for explaining the operation of the apparatus of FIG.

[Explanation of symbols]

101 ... Parameter decoding unit 102, 102 '... Audio signal reproducing unit 103, 103 '... Post filter 111 ... Pitch component emphasis filter 112 ... Spectral envelope enhancement filter 113 ... Compensation filter 114 ... Gain adjusting section 121 ... Adaptive filter 122 ... Filter coefficient calculation unit 123 ... First parameter calculation unit 124 ... Second parameter calculator 201 ... Synthesis filter information generation unit 202 ... Drive signal generator 203 ... First synthesis filter 204 ... Pitch component emphasis filter 205 ... Pitch tilt compensation filter 206 ... First LPC analysis unit 207 ... Second LPC analysis unit 208 ... Second synthesis filter 403 ... Post filter 410 ... Filter processing unit 414 ... Gain adjusting unit 415 ... Gain calculation unit 416 ... Gain multiplication unit 417 ... Addition unit 420-423 ... Calculation unit 424 ... Positive gain output section 425 ... Gain determination unit 426 ... Gain determination unit 2103 ... Post filter 2121 ... Adaptive filter 2122 ... First filter coefficient calculation unit 2123 ... Second filter coefficient calculation unit 2200 ... Parameter calculation unit 2201 ... Parameter calculation unit 2202 ... Zero filter 2203 ... Polar filter

Front page continued (72) Inventor Masami Akamine 1 Komukai Toshiba-cho, Saiwai-ku, Kawasaki-shi, Kanagawa Toshiba Research & Development Center Co., Ltd. (72) Inventor Emperor Amata 1 Komukai-shiba-cho, Kawasaki-shi, Kanagawa Corporate Toshiba Research and Development Center (56) Reference JP-A-3-245195 (JP, A) JP-A-2-82710 (JP, A) Special Table 2-502135 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 19/00 G10L 19/06

Claims (10)

(57) [Claims] 1. A first filter having a pole-zero type transfer function represented by A (z) / B (z) for an audio signal, and a second filter for compensating for the characteristics of the first filter. In the method of adjusting the spectral shape of an audio signal by passing it through a filter, the second filter has a transfer function in the z-transform domain of (1-
μ _z z ⁻¹ ) / (1−μ _p z ⁻¹ ) (where μ _z , μ
_p is at least a first-order pole-zero type transfer function represented by mutually independent filter coefficients each having an absolute value smaller than 1. _p . 2. A first filter having a pole-zero type transfer function represented by A (z) / B (z) for an audio signal, and a second filter for compensating for the characteristic of the first filter. In the method of adjusting the spectral shape of an audio signal by passing it through a filter, two parameters used for the second filter are A
(Z) and B (z) are calculated separately, and the second filter has a transfer function of (1-μ _z
z ⁻¹ ) / (1−μ _p z ⁻¹ ) (where μ _z and μ _p are mutually independent filter coefficients having absolute values smaller than 1) and the first-order pole-zero type transmission A spectrum shape adjusting method for an audio signal, comprising at least a function. 3. The first parameter generating means for predicting a characteristic of A (z) or 1 / A (z) and generating a prediction coefficient thereof as the first parameter; and B ( z) or 1 / B (z), and a second predicting coefficient is generated as the second parameter.
3. The method of adjusting the spectrum shape of an audio signal according to claim 2, further comprising: 4. The μ _p is obtained from the A (z), and the B
2. The μ _z is obtained from (z).
Alternatively, the method of adjusting the spectrum shape of an audio signal according to the item 2. 5. When the first autocorrelation coefficient obtained from the parameter of A (z) is smaller than a threshold value close to 0 using the weighting coefficient of the relation of C1 <C3 <C0, the first self The correlation coefficient is weighted with the weighting coefficient C0, and when the first autocorrelation coefficient is larger than the threshold value, the first autocorrelation coefficient is weighted with the weighting coefficient C1. 5. The μ _p is obtained, and the μ _z is obtained from a value obtained by weighting the second autocorrelation coefficient obtained from the parameter of B (z) with a weighting coefficient C3. Shape adjustment method of the voice signal of the. 6. A method of adjusting a spectral shape of an audio signal by performing a predetermined filtering process on the audio signal, wherein the audio signal is multiplied to compensate for a power change of the audio signal due to the filtering process. When adjusting the gain
The gain for compensating the power change of the audio signal due to the filter processing is calculated by a predetermined method, the positive / negative judgment of the gain is performed before multiplying the audio signal, and if the determination result is negative, the predetermined method is used. A method for adjusting a spectral shape of an audio signal, characterized by replacing a gain with a given non-negative value. 7. A method of adjusting a spectral shape of an audio signal by performing a predetermined filtering process on the audio signal, wherein the audio signal is multiplied to compensate for a power change of the audio signal due to the filtering process. When adjusting the gain
A gain for compensating the power change of the audio signal due to the filtering process is calculated by a predetermined method, and whether the gain is positive or negative is determined before being multiplied to the audio signal. If the determination result is negative, 0 or more, 1 A method for adjusting the spectral shape of an audio signal, characterized in that the gain is replaced with a non-negative value less than. 8. A first filter having a pole-zero type transfer function for enhancing a spectrum envelope of an audio signal and a second filter for compensating for a spectrum inclination by the first filter are cascade-connected. A step of independently obtaining two filter coefficients used in the second filter for compensating the slope of the spectrum from the pole-zero type transfer function, and the pole-zero type according to the obtained filter coefficient. Compensating the slope of the spectrum corresponding to each transfer function, the second filter has a transfer function in the z-transform domain of (1-μ
_z z ^-1 ) / (1-μ _p z ^-1 ), where μ _z and μ _p are first-order pole-zero type transmissions represented by mutually independent filter coefficients having absolute values smaller than 1. A method for adjusting a spectral shape of an audio signal, the method including at least a function. 9. A first filter having a pole-zero type transfer function for emphasizing a spectrum envelope of an audio signal, and an input from the first filter for compensating for a slope of a spectrum by the first filter. From the pole-zero type transfer function, 2
A filter for compensating for the slope of the spectrum corresponding to each of the pole-zero type transfer functions by filtering the audio signal output from the first filter in accordance with a calculation section for independently calculating two filter coefficients and the calculated filter coefficient. And a second filter having a section, the second filter has a transfer function of (1-μ _z
z ⁻¹ ) / (1−μ _p z ⁻¹ ), where μ _z and μ _p are first-order pole-zero type transfer functions represented by mutually independent filter coefficients having absolute values smaller than 1. A spectrum shape adjusting device for an audio signal, comprising at least: 10. A synthesis filter analyzer for analyzing an input voice signal and outputting synthesis filter data, a filter data calculator for calculating weighting filter data based on the synthesis filter data from the synthesis filter analyzer, and a weighting filter. A weighting filter for filtering an input audio signal based on the data, the weighting filter having a pole-zero transfer function constituted by using the weighting filter data, and compensating for the slope of the spelltle by the first filter. A second filter having a transfer function for
The transfer function of the conversion region is (1-μ _z z ^-1 ) / (1-μ _p z
^-1 ), where μ _z and μ _p include at least a first-order pole-zero type transfer function represented by mutually independent filter coefficients having absolute values smaller than 1, .