JP2001117573A - Audio spectrum enhancement method / apparatus and audio decoding apparatus - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To prevent an inappropriate spectrum tilt from occurring, and not to shift a convex part of a spectrum or emphasize a concave part.
Enables an ideal voice spectrum. After calculating an LPC coefficient which is an approximate amplitude spectrum of an audio signal by an LPC coefficient calculating unit, a convex frequency / recess frequency determining unit calculates a convex frequency and a concave frequency of the approximate amplitude spectrum, From these convex portion frequency and concave portion frequency, the convex portion band / recess band determining portion 104 is used.
The convex band and the concave band including the convex band frequency and the concave band frequency, respectively, are determined, and the characteristic that the amplitude spectrum of the frequency component included in the convex band is emphasized and the amplitude spectrum of the frequency component included in the concave band is attenuated. The filter having the filter configuration is configured by the filter configuration unit 106, and the characteristic of the filter is multiplied by the multiplier 109 to filter the audio signal, thereby performing spectrum enhancement.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a spectrum emphasizing method and apparatus for emphasizing a convex part of a spectrum of an audio signal and attenuating a concave part, and an audio decoding apparatus using the same.

[0002]

2. Description of the Related Art A technique for efficiently compressing and encoding a voice signal at a low bit rate is an important technique for effective use of radio waves and reduction of communication costs in mobile communications such as car telephones and in-company communications. It is. As a speech encoding method capable of producing high quality speech at a bit rate of 8 kbps or less,
The CELP (Code Excited Linear Prediction) method is known.

[0003] The CELP method is based on MR Schrodeder and BSAtal
“Code-Excited Linear Prediction (CELP) Hig
h-Quality Speech at Very low Bit Rates ”, Proc. IC
Since it was published in ASSP; 1985, pp.937-939 (Reference 1),
Attention has been paid to a method capable of synthesizing high-quality speech, and various studies have been made on improving quality and reducing the amount of calculation. However, at a low bit rate of 8 kbit / s or less, the quality of decoded speech is not yet sufficient.

[0004] Against this background, several techniques have been reported for performing post-processing on decoded speech to improve the perceptual quality of the decoded speech. For example, AT & T Bell Labs P.Kroo
n and BSAtal et al. “Quantization Procedures for
the Excitation in CELP Coders ”, Proc. ICASSP; 19
87, pp. 1649-1652 (Literature 2), a post-processing filter in which the characteristic is blunted by multiplying the LPC coefficient (linear prediction coefficient) sent to the decoder by a coefficient, and decoding is performed using this filter We report a method to obtain synthesized speech by filtering speech. If this post-processing filter is represented by the z-transform domain, it is as shown in equation (1).

[0005]

(Equation 1)

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

[0007]

(Equation 2)

However, an all-pole filter F1 as shown in equation (1)
In the case of (z), there is a problem that an improper spectrum inclination is included and the synthesized speech is muffled. Japanese Patent No. 2887286 (Reference 3) discloses a spectrum enhancement method for solving this problem. This document 3 proposes a method of cascade-connecting a pole-zero filter and a first-order zero-filter in consideration of spectral tilt correction. Expressing the transfer function of this filter in the z-transform domain gives equation (3).

[0009]

(Equation 3)

According to this spectrum emphasis filter,
Since the term A (z / γ) and the term (1-μz-1) in the equation (3) work to correct the inappropriate spectral tilt of the term A (z / β), the problem that the synthesized speech is muffled is It is reduced. However, depending on the characteristics of the signal input to the spectrum emphasis filter, the degree of the spectrum tilt correction may be excessive or insufficient. As a result, a low-frequency part is emphasized in a certain portion of the output signal of the spectrum emphasis filter, so that the sound is muffled. In a certain part, the high-frequency part is emphasized, so that the sound becomes too bright. For this reason, when listening to the output signal of the spectrum emphasizing filter, a wobble impression is given. Also, (3)
The position of the convex part of the spectrum may be shifted or the concave part may be emphasized due to the influence of the term A (z / γ) in the equation. As a result, there arises a problem that the output signal of the spectrum enhancement filter is deteriorated.

[0011]

First, problems of the method described in Reference 1 will be described. In this method, a spectrum emphasis filter is configured according to equation (1) using LPC coefficients obtained on the decoding side. FIG. 21 shows a short section of a decoded speech.
The LPC spectrum (solid line) of the synthesis filter H (z) composed of the LPC coefficient α (i) and the spectrum characteristic (dotted line) of the spectrum enhancement filter F1 (z) proposed in Document 1 are shown. 0.4 is used for β of F1 (z). Also, the synthesis filter H (z) is

[0012]

(Equation 4)

It can be expressed as

As can be seen from FIG. 21, the spectrum characteristic of the spectrum emphasizing filter F1 (z) is such that the spectrum is emphasized in the low frequency band and the spectrum is attenuated toward the high frequency band. Tend to be overemphasized. This tendency is observed in voiced parts that occupy many sections of the audio signal. As a result, there is a problem that the speech signal after the spectrum emphasis is heard muffled.

Similarly, the characteristic (dotted line) of the spectrum emphasis filter F2 (z) according to Document 2 is shown in FIG.
The LPC spectrum of (z) is represented by a solid line. Looking at the characteristics of the spectrum emphasis filter F2 (z), the problem that the low band is excessively emphasized compared to that of F1 (z) is reduced. This is because the numerator A (z / γ) and the term (1-μz−1) in the equation (3) work to correct an inappropriate spectral tilt. In this example, β = 0.8, γ = 0.5, and μ = 0.4.

[0016] For this reason, although the problem of the muffled sound as seen in Document 1 is reduced in Document 2, even in this example, there is a spectral gradient, and the tendency to emphasize low frequencies remains. Also, there is no guarantee that this filter will properly correct the spectral tilt in another section, but rather the problem of overemphasizing low frequencies or overemphasizing high frequencies in some sections may always occur. .

In addition to the problem of spectral tilt correction,
There is also a problem that the frequency of the spectrum maximum and the spectrum minimum deviate. In FIG. 22, the frequencies of the spectral maximum values of H (z) are f1 and f2, respectively, and the local values of the spectral characteristics of the spectrum emphasis filter F2 (z) are fe. Here, since f1 ≠ fe and f2 ≠ fe, the frequencies f1 and f2 of the spectrum maximum value move from the actual values.
In this case, f1 is shifted to a higher frequency and f2 is shifted to a lower frequency. Similarly, it is possible for the frequency of the spectral minima to shift as well. For example, in FIG. 22, the frequency f3 of the spectrum minimum value is shifted to a higher frequency. The positional relationship between the frequency of the spectrum maximum value and the minimum value is closely related to the phoneme information, that is, if the frequency of the spectrum maximum value and the minimum value moves,
This will lead to quality degradation.

Further, when the frequencies of the two spectral maximums are close to each other, the spectral minimum sandwiched between the spectral maximums may be emphasized. In FIG. 22, it can be seen that the spectrum emphasis filter F2 (z) works so as to emphasize at the frequency f4 of the spectrum minimum value. Such an event also destroys the phonemic information and causes quality deterioration.

An object of the present invention is to provide an ideal speech spectrum emphasizing method / apparatus which does not generate an inappropriate spectrum tilt and does not shift a convex portion of a spectrum or emphasize a concave portion, and to use the same. An object of the present invention is to provide an audio decoding device.

[0020]

In order to solve the above-mentioned problems, a voice spectrum emphasizing method according to the present invention comprises a method of forming a convex band and a concave band including a convex frequency and a concave frequency of an approximate amplitude spectrum of an audio signal. Determined, emphasizes the amplitude spectrum of the frequency component included in the convex band, configures a filter having a characteristic of attenuating the amplitude spectrum of the frequency component included in the concave band, and filters the audio signal by the filter. Features. The convex part band and the concave part band are typically bands having the convex part frequency and the concave part frequency as their respective center frequencies.

The speech spectrum emphasizing device according to the present invention comprises:
Means for determining the approximate shape of the amplitude spectrum of the audio signal, means for determining the convex and concave frequencies of the approximate amplitude spectrum, and the convex band and the concave portion including the convex and concave frequencies from the convex and concave frequencies, respectively. A means for determining a band, and a filter having characteristics of emphasizing the amplitude spectrum of the frequency component included in the convex band and attenuating the amplitude spectrum of the frequency component included in the concave band; And a means for filtering

More specifically, another voice spectrum emphasizing device according to the present invention comprises: means for obtaining a rough shape of an amplitude spectrum of a voice signal; means for calculating a convex portion frequency and a concave portion frequency of the rough amplitude spectrum; Means for determining the convex band and the concave band including the convex frequency and the concave frequency, respectively, and the amplitude spectrum of the frequency component included in the convex band is emphasized by multiplying by a predetermined convex magnification, and is included in the concave band. The amplitude spectrum of the frequency component is attenuated by multiplying by a predetermined concave magnification, and the amplitude spectrum of the frequency component not included in the convex band and the concave band is not less than the maximum value of the convex magnification and not less than the minimum value of the concave magnification. And a means for multiplying the audio signal by the set magnification, and filtering the audio signal by the filter.

As described above, according to the present invention, the convex portion frequency and the concave portion frequency are obtained from the approximate amplitude spectrum of the audio signal, and the convex portion for performing the spectrum emphasis and the spectrum attenuation so that the convex portion frequency and the concave portion frequency do not shift. Determine the magnification and the recess magnification. Specifically, the frequency components of the convex band (frequency components near the convex frequency) are emphasized at the same magnification, and the frequency components of the concave band (frequency components near the concave frequency) are similarly attenuated at the same magnification. .

By performing such processing, an inappropriate spectral tilt such that the low band is excessively emphasized and the high band is excessively emphasized does not occur in principle. Further, since the convex portion frequency and its vicinity are emphasized at the same magnification, and the concave portion frequency and its vicinity are attenuated at the same magnification, there is no problem that the convex portion frequency and the concave portion frequency shift.

Further, the magnification is determined so that the convex frequency is emphasized and the concave frequency is attenuated. Therefore, even when the two convex frequencies are close to each other, the concave frequency sandwiched between the convex frequencies is determined. Can be avoided.

In the present invention, when filtering the audio signal, the processing may be performed in either the frequency domain or the time domain. Performing filtering in the frequency domain makes it possible to accurately control the degree of spectrum emphasis and spectrum attenuation, and also has an affinity with the coding method in the frequency domain, such as MBE (Multi-band Excitation) coding. There is an advantage that the coding efficiency is high and the coding can be easily applied. When the filtering process is performed in the time domain, the process of converting the audio signal into the frequency domain can be omitted, and the amount of calculation can be reduced.

As an outline of the amplitude spectrum, for example, LPC
A spectrum is determined. In this case, from the LPC spectrum which is the approximate shape of the amplitude spectrum, it is possible to obtain the accurate convex frequency and concave frequency with a relatively small amount of calculation, which can contribute to quality improvement.

In the present invention, it is desirable that at least one of the width and the frequency position of the convex band including the convex frequency of the amplitude spectrum is determined by the positional relationship between the convex frequency and the concave frequencies located on both sides thereof. This makes it possible to adaptively determine the convex frequency to be emphasized and the width or frequency position of the convex band including the convex frequency for each convex frequency, thereby contributing to quality improvement.

In the present invention, it is preferable to determine the magnification of the convex portion and the magnification of the concave portion based on the approximate shape of the amplitude spectrum. In this case, it is possible to contribute to quality improvement by adaptively controlling the convex magnification and the concave magnification according to the magnitude of the amplitude spectrum of the frequency component included in the convex band to be emphasized and the concave band to be attenuated. Can be.

Furthermore, the speech decoding apparatus according to the present invention
The decoded speech signal and the parameter from the speech decoding unit that decodes the encoded data of the speech signal and outputs parameters such as LPC coefficients including information on the decoded speech signal and at least the amplitude spectrum of the speech signal, and the speech spectrum emphasis device described above. Is input to the spectrum emphasizing unit.
In this case, the spectrum emphasizing unit obtains a spectrum outline from parameters such as LPC coefficients and performs the same filtering process on the decoded speech signal to perform spectrum emphasis.

[0031]

DESCRIPTION OF THE PREFERRED EMBODIMENTS [First Embodiment] The first embodiment of the present invention
As an embodiment of the present invention, an example in which an input signal in a time domain is subjected to a filtering process using a spectrum enhancement filter in a frequency domain will be described.

FIG. 1 shows a spectrum emphasizing device according to a first embodiment of the present invention.
PC spectrum calculating section 102, convex section / concave section frequency determining section 103, convex section / concave section band determining section 104, convex section / concave section determining section 105, filter configuration section 106,
Audio input terminal 107, time-frequency converter 108, multiplier 109, frequency-time converter 110, gain calculator 1
11, a multiplier 112 and an audio output terminal 113.

First, the processing procedure of this embodiment will be described with reference to the flowchart shown in FIG. Input terminal 10
The LPC coefficient input from 1 is calculated by the LPC spectrum calculator 10.
2 and the projection magnification / recess magnification determination unit 105. LP
The C spectrum calculation unit 102 calculates an LPC spectrum using the LPC coefficient (step S1001), and then the convex frequency / recess frequency determining unit 103 determines the convex frequency and the concave frequency using the LPC spectrum (step S1001). S1002). Further, the convex band / concave band determination unit 104 determines the convex band and the concave band using the convex frequency and the concave frequency (step S1003).

In the convex / concave magnification determining unit 105, LP
The magnification of the convex portion and the magnification of the concave portion are obtained using the C coefficient (step S1004). In the filter configuration unit 106, the convex band /
A spectrum emphasis filter is configured using the convex band and concave band determined by the concave band determining unit 104 and the convex magnification and the concave magnification determined by the convex magnification / concave magnification determining unit 105 (step S1005).

On the other hand, the audio signal input from the input terminal 107 is converted into a frequency domain signal by the time-frequency converter 108 (step S1006). Spectrum of audio signal converted to this frequency domain and filter configuration unit 1
The filter 109 is multiplied by the filter characteristic coefficient of the spectrum emphasizing filter constituted by the multiplier 106 (step S100).
7), the output signal of the multiplier 109 is converted into a time-domain signal by the frequency-time conversion unit 110 (step S1008).

The gain calculator 111 calculates a correction gain so that the magnitude of the output signal of the frequency-time converter 110 matches the magnitude of the signal input from the input terminal 107 (step S1009). Is a multiplier 11
The output signal of the frequency-time conversion unit 110 is multiplied by 2 (step S1010). Then, the corrected audio signal is output from the output terminal 113.

Next, the operation of this embodiment will be described in more detail. The LPC coefficient input from the input terminal 101 is {α
(i); i = 1 to NP} (where NP represents the order of the LPC coefficient), the LPC spectrum X (n) calculated by the LPC spectrum calculation unit 102 using the LPC coefficient is expressed by the following equation (4). ) Can be obtained by replacing the filter given by z) with z = exp (j2πn / N). That is, the LPC spectrum X (n) is

[0038]

(Equation 5)

## EQU1 ## When the sampling period of the input audio signal is Fs [Hz], the spectral resolution of the LPC spectrum X (n) is Fs / N [Hz]. Assuming that the sampling period is Fs = 8000 [Hz] and N = 1000, the LPC spectrum X (n) is represented by a resolution of 8 [Hz].

In this embodiment, a method using an LPC spectrum as a means for obtaining a spectrum outline will be described. Alternatively, a spectrum outline may be obtained using an LPC cepstrum and an FFT cepstrum.

Next, the convex frequency / recess frequency determining section 10
In 3, the frequency at which the LPC spectrum X (n) has the maximum value and the minimum value, that is, the convex portion frequency and the concave portion frequency are obtained. The LPC spectrum X (n) takes a maximum value or a minimum value when a value dX (n) / dn obtained by differentiating the LPC spectrum X (n) by n becomes 0. In practice, when Equation (6) is satisfied and Equation (7) is satisfied, a local maximum exists at the frequency n, and Equation (6) is satisfied and Equation (6) is satisfied.
If (8) is satisfied, it is considered that a local minimum exists.

[0042]

(Equation 6)

The frequency n when the LPC spectrum X (n) takes a maximum value is expressed as a convex frequency {pk (j); j = 1 to NPK}, and the frequency n when the LPC spectrum takes a minimum value is expressed as a concave frequency {vy {k}; k = 1 to
NVY}. This is shown in FIG. In FIG. 3, 500H
z, 1000Hz, 2500Hz, 3200Hz, the spectral maximum is 800
There are spectral minima at Hz, 2000Hz and 3000Hz. Therefore, NPK = 4, NVY = 3, and the convex part frequency {pk (j); j =
1 to 4} and concave frequency {vy {k}; k = 1 to 3} are p
k (1) = 500, pk (2) = 1000, pk (3) = 2500, pk (4) = 3200, vy
(1) = 800, vy (2) = 2000, vy (3) = 3000.

Next, the convex band / concave band determination section 104 determines the convex frequency {pk (j); j = 1 to NPK} and the concave frequency {vy
{k}; k = 1 to NVY} and convex band {Bpk (j); k = 1 to
NPK} and concave band {Bvy (k); k = 1 to NVY}.

In the outline of the amplitude spectrum of the voice, the outline of the spectrum in the vicinity of the convex frequency has a different shape.
Similarly, the spectrum shapes near the concave portion frequency also have different shapes. For example, the convex portion frequency (pk (1), pk
(2), pk (3), pk (4)) and the concave frequency (vy (1), vy
Paying attention to the spectral outlines near (2) and vy (3), it can be seen that they have different shapes. Therefore, it is desirable to determine the convex zone and the concave zone according to these shapes.

An example in which the convex band and the concave band are adapted to the shape of the approximate spectrum near the convex frequency and the concave frequency will be described in a second embodiment which will be described later, and will be described in this embodiment. For the sake of simplicity, a method for determining the convex zone and the concave zone with a predetermined fixed width (bandwidth) will be described.

FIG. 4 schematically shows the relationship between the convex band {Bpk (j); j = 1 to NPK} and the concave band {Bvy (k); k = 1 to NVY}. In FIG. 4, the convex band is set so as to have a total bandwidth of 240 Hz around the convex frequency. Similarly, the recessed band is the whole around the recessed frequency.
It is set to have a bandwidth of 120Hz. The convex band / concave band determining unit 104 determines the convex band and the concave band in this way and gives the information to the filter configuration unit 106.

FIG. 5 shows the case where the set value of the convex band is 500 Hz and the set value of the concave band is 100 Hz. In this case, as shown in FIG. 5, an overlapping portion appears between the convex zone and the concave zone. In such a case, the convex band and the concave band are set so as to give priority to the convex band. FIG.
In the above, since the concave band corresponding to the concave frequency of 800 Hz and 3000 Hz overlaps with the convex band, the overlapping portion of the concave band is removed. On the other hand, convex part frequency 500Hz, 1000Hz
Are overlapped with each other. In such a case, these are regarded as one convex band. as a result,
The convex band {Bpk (j); j = 1 to NPK} and the concave band {Bvy (k); k =
1 to NVY} are as shown in FIG. Where NPK =
3. Corrected to NVY = 1.

The convex magnification / concave magnification determining unit 105 determines at least one of the convex magnification and the concave magnification using the LPC coefficient given from the input terminal 101. These magnifications are used to control the strength of the spectral enhancement filter. The basic idea is that if the difference between the amplitude of the spectral maximum and the amplitude of the spectral minima is large in the outline of the amplitude spectrum of the input audio signal, it is considered that the formant appears clearly, and the spectral emphasis is strongly applied. In this case, it is sufficient to consider that the formant hardly appears, and to weaken the spectrum.

In the present embodiment, the prediction gain of the synthesis filter represented by the equation (5) is used to correspond to the appearance of the formant. That is, when the prediction gain of the synthesis filter is large, the spectrum emphasis filter is applied strongly, and when the prediction gain is small, the spectrum emphasis filter is applied weakly.

More specifically, the filter gain of the synthesis filter is estimated by equation (9), and the magnification of the convex portion and the magnification of the concave portion are determined according to the estimated value. For the sake of simplicity, in the present embodiment, a case will be described in which the concave portion magnification is fixed to 1.0 and the convex portion magnification is obtained from the prediction gain of the synthesis filter. When the predicted gain of the synthesis filter is expressed in decibels,

[0052]

(Equation 7)

It is estimated that Here, ref (i) represents a PARCOR coefficient, and can be obtained from the LPC coefficient α (i) by a known method. Next, the convex portion magnification is determined according to Expression (10) using the prediction gain.

[0054]

(Equation 8)

Here, TA and TB represent constants for determining the magnification of the convex portion, and MIN () represents a function for outputting the minimum value. In this embodiment, TA = 1.0 and TB = 3.0. The convex magnification {Gpk (j); j = 1 to NPK} and concave magnification obtained in this way
{Gvy (k); k = 1 to NVY} is provided to the filter configuration unit 106.

In this embodiment, for simplicity, a method of setting all the convex magnifications {Gpk (j); j = 1 to NPK} to the same value according to the equations (10) and (11) will be described. However, a method of adaptively setting the magnification of each convex portion may be used. For example, based on the average value of the approximate amplitude spectrum included in each of the convex bands {Bpk (j); j = 1 to NPK}, the convex magnification {Gpk
(j); a method of setting j = 1 to NPK} is conceivable. Also,
In addition to the present embodiment, it is also possible to use a method of fixing the convex magnification to 1.0 and setting the concave magnification in accordance with the result of Expression (9). Similarly, it is also possible to use a method of setting both the magnification of the convex portion and the magnification of the concave portion.

In the filter configuration section 106, the convex band {Bpk (j); j = 1 to NPK} from the convex band / recess band determining section 104.
And information of the concave band {Bvy (k); k = 1 to NVY}, and the convex magnification {Gpk (j); j
= 1 to NPK} and the concave portion magnification {Gvy (k); k = 1 to NVY} to construct a spectrum emphasis filter. Here, the case where the convex band and the concave band shown in FIG. 4 are used, and all the convex magnifications are Gpk times and all the concave magnifications are Gvy times will be described. The schematic diagram is shown in FIG.

In FIG. 6, the convex band {Bpk (j); j = 1 to
NPK}, the characteristic of the spectrum emphasis filter is determined so that the amplitude spectrum of the frequency component becomes Gpk times. Similarly, the characteristics of the spectrum emphasizing filter are determined such that the amplitude spectrum of the frequency component included in the concave band {Bvy (k); k = 1 to NVY} is multiplied by Gvy.

Next, a magnification corresponding to a frequency that does not belong to any of the convex band and the concave band is determined. It is desirable that the magnification corresponding to this frequency be in the range of not more than the maximum value of the projection magnification and not less than the minimum value of the depression magnification. Furthermore, it is desirable to determine based on two magnifications defined by a convex band or a concave band near the low frequency side and a convex band or a concave band near the high frequency side to the current frequency of interest. .

This will be specifically described with reference to FIG. In FIG. 6, let ft be the current frequency of interest that does not belong to any of the convex band and the concave band and whose magnification is not yet determined. The magnification for the component of the frequency ft is a magnification Gpk corresponding to the convex band Bpk (2) existing on the low frequency side of ft, and a magnification Gvy corresponding to the concave band Bvy (2) existing on the high frequency side of ft. Set using. As a specific example, a method of determining the magnification by linear interpolation can be considered. In FIG. 6, when the highest frequency of the convex band Bpk (2) is fpk and the lowest frequency of the concave band Bvy (2) is fvy, the magnification Gft for the frequency ft component is

[0061]

(Equation 9)

Is expressed as follows. The magnification for other frequency components for which the magnification has not been determined can be determined in the same manner.

The frequency in FIG. 6 is 0.0 to fl [Hz].
The magnification for the frequency component belonging to the range of (fl is the lowest frequency of Bpk (1)) cannot be obtained because the magnification at the frequency 0.0 is not determined. In this case, the frequency 0.0
It is necessary to obtain the magnification at assuming that it is between Gvy and Gpk. In the present embodiment, the magnification at the frequency 0.0 is the magnification corresponding to the band to which fl belongs and the magnification at the opposite pole (Gvy in this case)
Suppose that Similarly, frequency fh ~ 4000 [Hz] (f
h is the maximum frequency of Bpk (4)).
The magnification corresponding to the band to which fh belongs and the magnification at the opposite pole (in this case, G
vy) and find the magnification. The magnification vector {G (n); n = 0 to N / 2} determined according to the above procedure is
As shown in FIG.

As another method of determining the magnification for the frequency component that does not belong to either the convex band or the concave band in the filter configuration unit 106, there is a method of limiting the magnification so that the magnification does not change more than a predetermined threshold value. is there.
For example, the frequencies f1 to f2, f3 to f4,
The magnification between f5 and f6 changes rapidly. In such a section, excessive spectrum emphasis and spectrum attenuation are performed at adjacent frequencies, which may cause quality deterioration.

In order to avoid this problem, it is effective to limit the magnification so that the magnification does not change more than a predetermined threshold value. As a specific restriction method, there is a method for preventing the absolute value of the slope of the interpolation function at the time of linear interpolation from being equal to or larger than a threshold value. Using this method, when the current frequency of interest is ft, the magnification Gft for this frequency ft is

[0066]

(Equation 10)

Can be obtained. Here, Tsl is a threshold value for limiting the magnification, and sign () is a function that returns a sign.

FIG. 8 shows the state of the magnification vector {G (n); n = 0 to N / 2} in the case where the above-mentioned restriction is provided. In FIG. 8, the same projection band, recess band, projection magnification, and depression magnification as those in FIG. 7 are used. As can be seen from FIG. 8, there is no portion where the magnification changes abruptly at adjacent frequencies, and the above-described problem can be avoided.

Next, an audio signal s (i) is given from the input terminal 107. It is assumed that the input audio signal and the LPC coefficient input from the input terminal 101 have a temporal correspondence, that is, are synchronized. In the time-frequency conversion unit 108,
The audio signal input from the input terminal 107 is converted into a signal in the frequency domain by a known technique such as DFT or DCT. At that time, if necessary, window hanging is also performed. The spectrum obtained by performing time-frequency conversion on the input signal is expressed as (S (n); n = 0 to N /
2}.

Next, the spectrum {S (n); n = 0 to N / 2} of the audio signal input to the multiplier 109 and the scale vector
{G (n); n = 0 to N / 2} is multiplied according to the following equation. However,
The spectrum after the multiplication is set to {U (n); n = 0 to N / 2}.

[0071]

[Equation 11]

This processing means that the spectrum of the input audio signal is emphasized and attenuated in the frequency domain by the magnification vector.

Next, the spectrum {U (n); n = 0 to N / 2} calculated by the multiplier 109 is converted by the frequency-time conversion unit 108 into a signal in the time domain. This conversion is defined as an inverse conversion of the time-frequency conversion unit 108. The signal in the time domain obtained by the time-frequency conversion unit 108 is represented by u
(i).

Next, the gain calculating section 111 calculates a gain for performing gain correction of the signal u (i) after spectrum enhancement. The gain calculator 111 calculates a correction gain gv according to the equation (15) using the input audio signal s (i) and the signal u (i) after spectrum enhancement.

[0075]

(Equation 12)

The multiplier 112 multiplies the signal u (i) after the spectrum emphasis by the correction gain gv to obtain the signal v after the gain correction.
Generate (i). The signal v (i) after gain correction is obtained according to the following equation. The signal v (i) after the gain correction is output from the output terminal 113. In this way, an audio signal whose spectrum has been emphasized can be obtained.

[0077]

(Equation 13)

In the present embodiment, an example has been described in which the gain correction is performed using the signals in the time domain, but the gain correction may be performed using the signals in the frequency domain. Specifically, the correction gain may be calculated such that the power of the output signal of the multiplier 109 matches the power of the output signal of the time-frequency conversion unit 108.

[Second Embodiment] In the outline of the amplitude spectrum of the sound, the outline of the convex frequency and the spectrum in the vicinity thereof have different shapes depending on the sharpness and the width of the peak. As an example, the convex portion frequency (pk (1), pk
(2), pk (3), pk (4))
It can be seen that each has a different shape. Therefore, it is desirable to determine the convex band according to the shape of these spectral outlines, as described in the description of the first embodiment.

Therefore, as a second embodiment of the present invention,
An embodiment for determining a convex band suitable for the shape of the spectrum shape near the convex frequency will be described with reference to FIG.
FIG. 9 shows a state of one convex frequency pk (j) and two concave frequencies vy (k) and vy (k + 1) located at both ends thereof.

First, a frequency FL located at an internal dividing point based on a predetermined ratio between the convex portion frequency pk (j) and the concave portion frequency vy (k) located on the left side is obtained. Similarly, the convex frequency
A frequency FR located at an internally dividing point based on a predetermined ratio between pk (j) and the concave portion frequency vy (k + 1) located on the right side is obtained. Then, the band from the frequency FL to FR is changed to the convex portion frequency pk.
(j) The convex band Bpk (j) with the sum center as the center. In FIG. 9, the ratio at the time of obtaining the internally dividing point is set to 0.5. In this method, a limit may be set to a predetermined length or less so that the convex band Bpk (j) does not become too large.

Further, as shown in FIG.
Concave frequencies vy (k), vy (k + 1) located on both sides of (j)
If any one of these (vy (k + 1) in this example) is significantly farther away than the other (vy (k) in this example), the convex band Bpk (j) may be extremely displaced. There is.

To avoid this problem, the projection frequency pk
A method of limiting the convex band using the distance between (j) and the frequencies FL and FR located at the inner dividing point is considered. In particular,
The distance between the convex frequency pk (j) and FL is DL, and the convex frequency pk (j) is
Assuming that the distance from FR is DR, the distances DL and DR are expressed as follows.

[0084]

[Equation 14]

Next, the size of DL is a constant C of DR (C> 1.0
) It is determined whether the number exceeds DR or the size of DR exceeds C times DL. If so, DL or DR is limited. In particular,

[0086]

(Equation 15)

And the frequencies FL ′ and FR ′ corresponding to DL ′ and DR ′
And a band from FL ′ to FR ′ is defined as a convex band Bpk (j) of the convex frequency pk (j). FIG. 10 shows this state. Here, the constant C is set to 2.0. Also in this method, a limit may be set to a predetermined length or less so that the convex band Bpk (j) does not become too large.

[Third Embodiment] FIG. 11 shows the configuration of a speech spectrum enhancing apparatus according to a third embodiment of the present invention. In FIG. 11, components having the same names as those in FIG. 1 have the same functions as those in FIG. 1, and a description thereof will be omitted.

The first embodiment is different from the first embodiment in that the spectrum emphasis filter is configured in the frequency domain, whereas the second embodiment is configured in the time domain. The time-frequency conversion unit 108 in FIG. The frequency-time conversion unit 110 has also been removed. FIG. 12 is a flowchart showing the flow of the processing of this embodiment. The processing of steps S2001 to S2004 is the same as steps S1001 to S1004 of FIG.

In this embodiment, in the filter configuration section 206, first, the convex band and the concave band determined by the convex band / recess band determining section 204, and the convex section determined by the convex magnification / recess magnification determining section 205. After determining the characteristic (or magnification vector) {G (n); n = 0 to N / 2} of the spectrum emphasizing filter as shown in FIG. 7 and FIG. A time-domain filter is configured (step S2005).

More specifically, first, a power spectrum of the characteristic of a desired spectrum emphasizing filter is obtained, and this power spectrum is subjected to inverse Fourier transform to obtain an autocorrelation function. Then, an LPC coefficient {ap (m); m = 1 to M} is obtained. The spectrum emphasis filter in the time domain is configured using the LPC coefficient {ap (m); m = 1 to M} obtained as described above as the characteristic of the all-pole filter, and the filtering unit 2 is configured using this filter.
10, filtering is performed to enhance the spectrum of the input audio signal (step S2006), and then a correction gain is calculated by the gain calculator 211 (step S2006).
2007), the multiplier 212 multiplies the audio signal after spectrum enhancement by the multiplier 212 (step S200).
8).

Further, based on the autocorrelation function described above, a pole-zero filter is generated according to a known method so as to be close to a desired characteristic to constitute a time-domain spectrum emphasis filter.
Filtering can be performed by the filtering unit 210 using this filter, and the spectrum of the input audio signal can be emphasized. As such a filter, a Modified Yule Walker filter and the like are known.

[Fourth Embodiment] FIG. 13 shows a configuration of a speech spectrum emphasizing apparatus according to a fourth embodiment of the present invention. In FIG. 13, components having the same names as those in FIG. 1 have the same functions as those in FIG.

In the first embodiment, the LPC spectrum obtained from the LPC coefficient is used as the approximate shape of the amplitude spectrum.
8 in that the amplitude spectrum outline calculation unit 302 obtains the amplitude spectrum outline using the signal in the frequency domain of the audio signal calculated in step S8. Therefore, in the present embodiment, there is no need to provide the LPC coefficient as an input. FIG. 14 is a flowchart showing the flow of processing according to the present embodiment. The processing in steps S3003 to S3010 is the same as that in steps S1003 to S1010 in FIG.

That is, the audio signal input from the input terminal 307 is first converted into a signal in the frequency domain by the time-frequency conversion unit 308 (step S3001), and the spectrum {S (n); n = 0 to N / 2}, the amplitude spectrum outline calculation unit 302 calculates the amplitude spectrum outline (step S3002). As a specific calculation method of the amplitude spectrum outline, for example, a spectrum {S (n); n = 0 to
N / 2}. Moving average spectrum {Sa
(n); n = 0 to N / 2} can be calculated as follows using the spectrum {S (n); n = 0 to N / 2}.

[0096]

(Equation 16)

M represents the window length of the moving average. The moving average spectrum obtained in this manner is provided to the convex / concave frequency determining unit 303.

In this embodiment, the convex / concave magnification determining unit 304 determines whether the spectrum {S (n); n = 0 to N / 2} of the input audio signal or the moving average spectrum {S
a (n); n = 0 to N / 2}, the convex portion magnification and the concave portion magnification are determined. As a specific method, for example, spectrum
A method using the power of {S (n); n = 0 to N / 2}, a method using the difference between the upper and lower limits of the moving average spectrum {Sa (n); n = 0 to N / 2}, etc. Conceivable.

[Fifth Embodiment] FIG. 15 shows the configuration of a speech spectrum enhancing apparatus according to a fifth embodiment of the present invention. In FIG. 15, components having the same names as those in FIG. 1 have the same functions as those in FIG. 1, and a description thereof will not be repeated. FIG. 16 is a flowchart showing the flow of the processing of this embodiment, and includes steps S4001 to S4005,
The processing in S4007 to S4009 is performed in steps S1001 to S100 in FIG.
5. Same as S1007 to S1009.

This embodiment is different from the first embodiment in that the signal supplied from the input terminal 407 is a signal in the frequency domain, for example, a voice spectrum. Therefore, the time-frequency conversion unit 108 in FIG. 1 is unnecessary, and the frequency spectrum of the audio signal is multiplied by the characteristic of the spectrum emphasis filter configured in step S4005 (step S4006).

The configuration of the present embodiment can be applied to a method of performing encoding in the frequency domain, for example, when a decoded signal is represented in the frequency domain as in MBE (Multi-band Excitation) encoding. In such a case, there is an advantage that the time-frequency conversion unit can be omitted.

Further, in the configuration of the present embodiment, the input gain spectrum and the multiplier 4
It is obtained by using the signal after spectrum emphasis which is the output of the signal 08. Further, in frequency-time conversion section 410, conversion is performed using the signal after gain correction, which is the output of multiplier 412, and this output signal is output from output terminal 413.

[Sixth Embodiment] FIG. 17 shows a configuration of a speech spectrum emphasizing apparatus according to a sixth embodiment of the present invention. In FIG. 17, components having the same names as those in FIGS. 1 and 15 have the same functions as those in FIGS. 1 and 15, and a description thereof will be omitted. FIG. 18 is a flowchart showing the flow of the processing of this embodiment. The processing in steps S5002 to S5009 is performed in steps S4002 to S40 in FIG.
Same as 09.

This embodiment is closest to the fifth embodiment described above, and differs from the fifth embodiment in that no LPC coefficient is given in this embodiment. In this case, similar to the fourth embodiment,
Using the voice spectrum given from the input terminal 507, the amplitude spectrum outline calculation unit 502 obtains an amplitude spectrum outline (step S5001).

[Seventh Embodiment] FIG. 19 shows an example in which a spectrum emphasizing device according to the present invention is applied to a speech decoding device in a speech encoding / decoding system as a seventh embodiment of the present invention. In the present embodiment, a case where the CELP method is used as a speech encoding / decoding system will be described, but the present invention is not limited to this. For example, MBE
It can also be applied to a method of performing speech coding in the frequency domain as described above. Also, in the present embodiment, a case will be described in which the spectrum emphasizing device shown in the first embodiment is used.
The present invention is not limited to this, and other embodiments can be applied. FIG. 2 showing a processing flow of the present embodiment.
The configuration and operation of the present embodiment will be described in combination with 0.

The speech decoding apparatus shown in FIG. 19 mainly comprises a speech decoding section 601 and a spectrum emphasizing section 602. Also, in FIG. 20, steps S6001 to S6007
Indicate the speech decoding process, and the processes indicated in steps S6008 to S6017 indicate the spectrum emphasis process.

From an input terminal 603, an encoded bit stream representing an audio signal compressed and encoded by an audio encoding unit (not shown) is input to the audio decoding unit 601. The input coded bit stream is separated and converted into an LPC coefficient index, an ACB vector index, an SCB vector index, and a gain index by the demultiplexer 604 (step S6001).

The LPC coefficient decoding section 605 decodes the LPC coefficient {αq (i); i = 1 to NP} based on the LPC coefficient index (step S6002). The decoded LPC coefficients are provided to synthesis filter 612 and spectrum enhancement section 602. ACB
The vector decoding unit 606 decodes the ACB vector using the ACB vector index (step S6003), and the SCB vector decoding unit 607 decodes the SCB vector using the SCB vector index (step S6004). Similarly, the gain decoding unit 608 decodes the ACB vector gain and the SCB vector gain using the gain index (step S6005).

The multiplier 610 multiplies the ACB vector by the ACB vector gain, and the multiplier 609 multiplies the SCB vector by S
Multiply the CB vector gain. These multipliers 610 and 6
The sound source is generated by adding the signals after the multiplication by 09 by the adder 611 (step S6006). Therefore, the sound source ex (n) is represented by the following equation.

[0110]

[Equation 17]

Here, p (n) represents the ACB vector, c (n) represents the SCB vector, gp represents the ACB vector gain, and gc represents the SCB vector gain.

The synthesized signal so is passed through a sound source ex (n) to a synthesis filter 612 composed of LPC coefficients {αq (i); i = 1 to NP}.
(n) is generated (step S6007). The composite signal so (n) is
It is obtained according to the following equation.

[0113]

(Equation 18)

The synthesized signal so (n) thus generated
Is given to the spectrum emphasizing unit 602. The spectrum enhancement unit 602 has the same configuration as the spectrum enhancement device described in the first embodiment shown in FIG.

In the voice spectrum emphasizing unit 602, the LP
In the C spectrum calculation unit 102, the LP including the information of the amplitude spectrum of the audio signal given from the audio decoding unit 601
Using the C coefficient {αq (i); i = 1 to NP}, an LPC spectrum that is an approximate amplitude spectrum is calculated as in the first embodiment (step S6008).

Next, the convex / concave frequency calculating section 10
3, the convex part frequency and the concave part frequency are detected using the LPC spectrum (step S6009), and the convex part band and the concave part band are determined by the convex part band / concave part band determining unit 104 based on the convex part frequency and the concave part frequency (step S6009). Step S6010).

Also, the convex / concave magnification determining unit 104 determines the convex and concave magnifications using the LPC coefficient {αq (i); i = 1 to NP} (step S6011). Then, the filter configuration unit 106 configures a spectrum enhancement filter based on the convex band and the concave band, and the convex magnification and the concave magnification (step S6012).

Next, the time-frequency converter 108 converts the synthesized signal so (n) into a signal in the frequency domain (step S60).
13) The multiplier 109 multiplies the frequency domain signal of the synthesized signal by the spectrum emphasis filter (step S601).
4) The frequency-time conversion unit 110 converts the multiplied signal into a time-domain signal (step S6015).

Next, the gain calculating section 111 calculates a correction gain using the composite signal so (n) and the output signal of the frequency-time converting section 110 (S6016). And the correction gain are multiplied by the multiplier 112 (step S6017) and output from the output terminal 613.

In this way, spectrum enhancement can be performed on the decoded speech signal output from speech decoding section 601 by speech spectrum enhancement section 602 according to the present invention.

[0121]

As described above, according to the present invention, the convex frequency and the concave frequency are obtained from the outline of the amplitude spectrum of the audio signal, and the amplitude spectrum is emphasized for the frequency component of the convex band. By attenuating the amplitude spectrum of frequency components, good spectrum emphasis is performed without in principle causing an inappropriate spectrum tilt such that low frequencies are excessively emphasized or high frequencies are excessively emphasized. be able to.

Further, by emphasizing the convex portion frequency and its vicinity at the same magnification and attenuating the concave portion frequency and its vicinity at the same magnification, there is no problem that the convex portion frequency and the concave portion frequency shift.

Further, the magnification is determined so that the convex frequency is emphasized and the concave frequency is attenuated. Therefore, even when the two convex frequencies are close to each other, the concave frequency sandwiched between the convex frequencies is determined. Can be avoided.

As described above, according to the present invention, it is possible to perform ideal spectrum enhancement without generating an inappropriate spectrum inclination and without shifting a convex part of a spectrum or emphasizing a concave part. .

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech spectrum emphasizing device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a processing procedure in the embodiment.

FIG. 3 is a view showing an example of a distribution of a convex portion frequency and a concave portion frequency for explaining the operation of the embodiment.

FIG. 4 is a diagram schematically showing an example of a relationship between a convex band and a concave band for explaining the operation of the embodiment.

FIG. 5 is a diagram schematically showing another example of the relationship between the convex band and the concave band for explaining the operation of the embodiment.

FIG. 6 is a diagram showing the relationship between the convex band and the concave band, and the convex magnification and the concave magnification for explaining the operation of the filter constituting unit in the embodiment.

FIG. 7 is a view showing an example of a method for determining a magnification of a convex portion and a magnification of a concave portion in the embodiment.

FIG. 8 is a view showing another example of the method of determining the magnification of the convex portion and the magnification of the concave portion in the embodiment.

FIG. 9 is a diagram illustrating a method of determining a convex band according to the second embodiment of the present invention.

FIG. 10 is a diagram showing another example of the method of determining the convex band according to the embodiment.

FIG. 11 is a block diagram showing a configuration of a speech spectrum emphasizing device according to a third embodiment of the present invention.

FIG. 12 is a flowchart showing a processing procedure according to the embodiment;

FIG. 13 is a block diagram showing a configuration of a speech spectrum emphasizing device according to a fourth embodiment of the present invention.

FIG. 14 is a flowchart showing a processing procedure according to the embodiment;

FIG. 15 is a block diagram showing a configuration of a speech spectrum emphasizing device according to a fifth embodiment of the present invention.

FIG. 16 is a flowchart showing a processing procedure according to the embodiment;

FIG. 17 is a block diagram showing a configuration of a speech spectrum emphasizing device according to a sixth embodiment of the present invention.

FIG. 18 is a flowchart showing a processing procedure according to the embodiment;

FIG. 19 is a block diagram showing a configuration of a speech decoding device according to a seventh embodiment of the present invention.

FIG. 20 is a flowchart showing a processing procedure according to the embodiment;

FIG. 21 is a diagram showing an LPC spectrum of a synthesis filter and a spectrum characteristic of a spectrum emphasis filter for explaining a first related art;

FIG. 22 is a view showing an LPC spectrum of a synthesis filter and a spectrum characteristic of a spectrum emphasis filter for explaining a second conventional technique.

[Explanation of symbols]

101, 201, 401... LPC coefficient input terminals 102, 202, 402, 502... LPC spectrum calculator 302... Amplitude spectrum shape calculator 103, 203, 303, 403, 503... 204, 304, 404, 504...
Concave band determining unit 105, 205, 305, 405, 505...
Concavity magnification determining units 106, 206, 306, 406, 506: Filter components 107, 207, 307: Audio input terminals 407, 507: Audio spectrum input terminals 108, 308: Time-frequency converters 109, 309: Multipliers 110 , 310, 410, 510 frequency-time conversion section 210 filtering section 111, 211, 311, 411, 511 gain calculation section 112, 212, 312, 412, 512 ... multipliers 113, 213, 313, 413, 513 , 613: audio output terminal 601: audio decoding unit 602: audio spectrum emphasis unit 603: audio encoded bit stream input terminal

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 5D045 CA01 CB01 5J064 AA01 BB03 BB08 BB12 BB13 BC02 BC09 BC11 BC27 BD01 CA01 CB13

Claims (9)

[Claims] An amplitude spectrum of an audio signal is determined by determining a convex band and a concave band including a convex frequency and a concave frequency, respectively, and the amplitude spectrum of a frequency component included in the convex band is emphasized. An audio spectrum emphasizing method, comprising: forming a filter having a characteristic of attenuating an amplitude spectrum of a contained frequency component, and filtering the audio signal by the filter. 2. A means for obtaining an approximate shape of an amplitude spectrum of an audio signal; a means for obtaining a convex frequency and a concave frequency of the approximate amplitude spectrum; and a convex frequency and a concave frequency from the convex frequency and the concave frequency, respectively. Means for determining a convex band and a concave band including, and a filter having a characteristic of enhancing an amplitude spectrum of a frequency component included in the convex band and attenuating an amplitude spectrum of a frequency component included in the concave band. hand,
Means for filtering the audio signal by the filter. 3. A means for obtaining an approximate shape of an amplitude spectrum of an audio signal; a means for obtaining a convex portion frequency and a concave portion frequency of the approximate amplitude spectrum; and a convex portion band and a concave portion band including the convex portion frequency and the concave portion frequency, respectively. Means for determining the amplitude spectrum of the frequency component included in the convex band by multiplying the amplitude spectrum of the frequency component by a predetermined convex magnification, and the amplitude spectrum of the frequency component included in the concave band by a predetermined concave section. Attenuated by multiplying by a magnification, for the amplitude spectrum of the frequency component not included in the convex band and the concave band, the magnification set to the maximum of the convex magnification or more and the minimum of the concave magnification or more. Means for constructing a multiplying filter, and means for filtering said audio signal with said filter. . 4. An audio spectrum emphasizing apparatus according to claim 2, wherein said filtering means performs said filtering processing in a frequency domain. 5. The filtering device according to claim 2, wherein the filtering unit performs the filtering process in a time domain.
Or the speech spectrum emphasizing device according to 3. 6. The means for determining the approximate shape of the amplitude spectrum comprises:
4. The speech spectrum emphasizing device according to claim 2, wherein an LPC spectrum is obtained as the amplitude spectrum outline. 7. The means for determining the convex band and the concave band determines at least one of the width and the frequency position of the convex band based on the positional relationship between the convex frequency and the concave frequencies located on both sides thereof. The speech spectrum emphasizing device according to claim 2 or 3, wherein: 8. An audio spectrum emphasizing apparatus according to claim 2, further comprising means for determining the magnification of the convex portion and the magnification of the concave portion based on the approximate shape of the amplitude spectrum. 9. An audio decoding unit for decoding encoded data of an audio signal and outputting a decoded audio signal and a parameter including at least information on an amplitude spectrum of the audio signal, a decoded audio signal from the audio decoding unit and the parameter And a spectrum emphasis unit configured by the speech spectrum emphasis device according to any one of claims 2 to 8, wherein the spectrum emphasis unit obtains the spectrum outline from the parameters, and obtains the decoded speech. An audio decoding device, which performs the filtering process on a signal.