JPH0990998A - Acoustic signal conversion decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the number of operation steps by reducing calculation and division of a power spectrum envelope in a decoder. SOLUTION: This method reproduces a signal by decoding a quantized sign after normalizing it by a square root power spectrum envelope, reproducing an inputted linear prediction parameter by a means 56, performing an operation to obtain a power spectrum concerning predetermined sparse frequency points from this linear prediction parameter, and a square root power spectrum envelope value is obtained by means 101. About an unoperated frequency point in the part where this change is comparatively large, the square root power spectrum envelope value can be determined by a means 103 by using the reproduction value of means 56, and about an unoperated frequency point in the part where this change is small, the square root power spectrum envelope value is obtained by interpolating them from their fore and aft values.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an acoustic signal conversion / decoding method for converting an acoustic signal such as voice or music into a frequency domain and decoding a highly efficient code into an acoustic signal.

[0002]

2. Description of the Related Art A conventional audio signal conversion coding method and its decoding method will be described with reference to FIG. The audio input signal sequence digitized from the input terminal 33 in the encoder 31 is input to the frame dividing means 34, and the input sequence of the past 2 × N samples is extracted for each N input samples, and the length is 2 ×.
It is generated in the input frame of N samples, and the windowing means 35 applies a time window to the input frame. A Hanning window is generally used as the window shape. The windowed input signal sequence is subjected to modified discrete cosine transform by the MDCT means 36 and transformed into a frequency domain signal of N samples.

Further, the windowed input signal sequence is subjected to linear prediction analysis by the linear prediction analysis means 37 to obtain a P-th order prediction coefficient. This linear prediction analysis is performed after obtaining the autocorrelation. The prediction coefficient is quantized by the quantizing means 38. As this quantization method, an LSP quantization method of converting a prediction coefficient into an LSP parameter for quantization, a method of converting a prediction coefficient into a k parameter and then performing quantization, or the like can be used. An index 39 indicating this quantized prediction coefficient is transmitted.

Further, the quantized prediction coefficient is obtained by calculating a power spectrum by a frequency outline calculation means 41 to obtain a frequency characteristic outline signal. Specifically, the quantized output of the quantizing means 38 is dequantized, and, for example, as shown in FIG. 4, the P + 1 dequantized prediction coefficients (α parameter) are followed by 2
An NFT power spectrum is calculated by FFT analysis (fast Fourier transform: discrete Fourier transform) of a sample sequence of length 2 × N formed by connecting × N−P−1 0s. Each point of the reciprocal of the i-th frequency characteristic outline starting from the 0-th is obtained by averaging, ie, interpolating, the square roots of the i + 1-th and i-th power spectra except i = N−1. The inverse of the N-1th frequency characteristic outline is obtained by taking the square root of the N-1th power spectrum.

Returning to the explanation of FIG. 3, in the normalizing means 42, each sample of the frequency domain signal from the MDCT means 36 is multiplied by each sample of the reciprocal of the frequency outline to be normalized and flattened. Is the residual signal. In the power normalization / gain quantization means 43, this residual signal is divided by the average value of its amplitude or the normalization gain which is the square root of the average value of the power to be normalized, and the normalized residual signal is obtained. Further, the normalized gain is quantized, and the index 44 indicating the quantized normalized gain is output.

The signal of the reciprocal of the frequency characteristic outline from the frequency outline calculating unit 41 is weighted by the weight calculating unit 45 as necessary.
The audibility control is applied to produce a weighted signal. The normalized residual quantizing means 46 adaptively weights the normalized residual signal from the means 43 using the weighting signal from the means 45. An index 47 indicating the vector value quantized by the quantizer 46 is output. As described above, from the encoder 31, the prediction coefficient quantization index 39,
The gain quantization index 44 and the residual quantization index 47 are output.

The decoder 32 to which these indexes 39, 44 and 47 are input decodes as follows as shown in FIG. That is, the predictive coefficient quantization index 39 is reproduced by dequantizing the corresponding quantized predictive coefficient by the reproducing means 56, and the dequantized predictive coefficient is reproduced by the frequency outline calculating means 57 in the same manner as the frequency outline calculating means 41. Is calculated the reciprocal of the frequency characteristic outline, that is, the reciprocal of the square root of the power spectrum envelope, while the index 4 input by the reproducing means 58 is calculated.
From 7, the quantized normalized residual signal is regenerated. Reproduction means 5
The normalized gain is reproduced from the index 44 input in 9. The quantized and normalized residual signal regenerated by the power denormalization means 61 is multiplied by the regenerated normalization gain and power denormalized to obtain a quantized residual signal. The quantized residual signal is divided by the inverse normalization means 62 from the frequency outline calculation means 57 for each corresponding sample by the inverse of the frequency outline, that is, the inverse of the square root of the power spectrum envelope, and inverse flattened. The inverse-flattened residual signal is subjected to Nth-order inverse modified discrete cosine transform in the inverse MDCT means 63 to be a time domain signal, and the windowing means 64 applies a time window to the time domain signal. Here, a Hanning window is used as the window shape. The windowed signal is added to the first half N samples of the frame having a length of 2.times.N samples and the second half N samples of the previous frame by the frame superimposing means 65 and output to the output terminal 66.

In the decoder 32, the index 39
The inverse quantized prediction coefficient is obtained from the obtained power spectrum, the power spectrum is obtained as shown in FIG. 4, the square root for each sample is obtained, and the reciprocal of this is obtained respectively. It requires a considerable amount of processing, which is an obstacle to real-time operation. From such a point, it is possible to simplify the processing on the decoding side by normalizing the signal (coefficient) converted into the frequency domain by the square root of the power spectrum and then quantizing the signal (Japanese Patent Application No. 7-3888). Proposed in. That is, as shown by attaching the same reference numerals to the parts corresponding to those in FIGS. 5 and 3, the output signal sequence of the windowing means 35 is an envelope representing the square root of the power spectrum envelope (hereinafter referred to as the square root power spectrum envelope). ) Is branched and supplied to a means 71 for modeling by using a linear prediction analysis. Means 7
1, the correlation function means 72 first obtains the autocorrelation function of the input signal up to N points in one frame. Next, this autocorrelation function at N points is added to this series by zeros at N points, or N points are made symmetric and substituted, and the real Fourier transform at 2N points is performed by the Fourier transform means 73. If the correlation function is obtained by the autocorrelation method, the real part after conversion is the power spectrum, and all take positive values except for errors in calculation accuracy.
It is well known that the power spectrum can be obtained by obtaining the autocorrelation of the input signal and performing Fourier transform on the autocorrelation.

The square root means 74 finds the square root of each point of this power spectrum. At this time, the imaginary part is set to all zeros (the imaginary part is originally all zeros when it is symmetric and substituted), and then the inverse Fourier transform is performed by the inverse Fourier transform means 75 to obtain the autocorrelation function corresponding to the square root power spectrum. Finally, linear prediction analysis is performed by the linear prediction analysis means 76 based on this autocorrelation function to obtain prediction parameters,
That is, the modeled by the linear prediction analysis expressing the square root of the power spectrum envelope is obtained. This is quantized by the prediction coefficient quantization means 38 to obtain the index 39. The index 39 is not obtained by linear prediction analysis of the input signal sequence, but is a quantized result of linear prediction analysis of the frequency characteristic of the input signal sequence, that is, the signal root corresponding to the square root of the power spectrum envelope.

The quantized output of the quantizing means 38 is inversely quantized by the inverse quantizing means 77, and the inversely quantized linear prediction coefficient is Fourier transformed by the Fourier transform and absolute value means 78,
The absolute value of the complex number of each sample is taken to obtain the reciprocal of the square root power spectrum envelope, and this is multiplied by the frequency domain signal from the MDCT means 36 by the multiplier 42 for each sample and normalized. The other processes are the same as in the case of FIG.

In contrast to such encoding, decoding may be performed as shown in FIG. In FIG. 6, the parts corresponding to those in FIG.
, The index 39 is dequantized to obtain a linear prediction coefficient. This linear prediction coefficient is obtained by performing a linear prediction analysis of the square root power spectrum, as is clear from the description of FIG. The inverse of the square root power spectrum envelope is obtained by performing Fourier transform with 82 and obtaining the absolute value thereof. The reciprocal of the inverse of the square root power spectrum envelope is taken by the reciprocal calculator 82 to obtain the square root power spectrum envelope, and this is multiplied. The reproduced residual signal from 61 is multiplied by a multiplier 84 to reproduce a frequency domain signal. Others are the same as in the case of FIG.

As described above, the decoder does not need to obtain the power spectrum envelope, does not need the square root operation, and the processing of the decoder is reduced accordingly.

[0013]

SUMMARY OF THE INVENTION It is an object of the present invention to provide an acoustic signal conversion decoding method capable of reducing the amount of calculation by reducing the calculation and division of the power spectrum envelope in the decoder.

[0014]

According to the invention of claim 1, a residual signal is obtained by dequantizing the first index in the input code, and the square root of the power spectrum envelope is obtained from the second index in the input code. In the acoustic signal conversion decoding method for obtaining the acoustic signal by denormalizing the residual signal to obtain the frequency domain signal by this, and converting the frequency domain signal into the time domain signal, the second index is In the first step, dequantization is performed to obtain a linear prediction parameter, in the second step, the square root value of the power spectrum envelope is obtained at predetermined sparse frequency positions in the linear prediction parameter.
From the square root value of the power spectrum envelope at
The square root value of the power spectrum envelope at the frequency position not obtained in the step is obtained by interpolation.

According to the second aspect of the present invention, the square root value of the power spectrum envelope of a part of the frequency positions which is not obtained in the second step is directly obtained from the linear prediction parameter in the fourth step. In the invention of claim 3, the interpolation in the third step is performed in the high frequency region, and the direct calculation in the fourth step is performed in the low frequency region.

In the invention of claim 4, the magnitude of the change rate of the square root of the power spectrum envelope obtained in the second step is judged in the fifth step, and the interpolation in the third step is performed at the frequency position judged to have the small change rate. Then, at the frequency position where the change rate is determined to be large, it is directly obtained in the fourth step. In the invention of claim 5, in the second step, the square root of the power spectrum envelope is obtained as a result of the calculation for obtaining the power spectrum envelope from the linear prediction parameter.

In the second aspect of the invention, in the second step, an operation for obtaining the power spectrum envelope is performed from the linear prediction parameter, and the square root of the operation result is obtained to obtain the square root of the power spectrum envelope.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an embodiment of the present invention in which parts corresponding to those in FIG. That is, in this example, as described with reference to FIG. 5, on the encoding side, the frequency domain signal is normalized by the square root of the power spectrum envelope of the input signal (square root power spectrum envelope) and then quantized code. This is a case where the present invention is applied to decrypt The prediction parameter reproducing means 56 dequantizes the prediction index 39 to obtain a linear prediction parameter. In this embodiment, the power spectrum calculation means 101 calculates the power spectrum envelope from the linear prediction parameter only at a predetermined frequency position. The calculation is performed, and as a result, the square root power spectrum envelope is obtained. The means 101 is, for example, means 10 for obtaining the reciprocal of the spectral envelope.
The reciprocal of the square root spectrum envelope is obtained by 2, and the reciprocal of the square root spectrum envelope is calculated by the reciprocal means 103. As the means 102, for example, when the predetermined frequency positions are at equal intervals, zeros are put after the linear prediction coefficient α obtained as a linear prediction parameter, and for example, N / 4 pieces (2N is in one frame) as a whole. Then, the inverse number of each power can be obtained for N / 8 frequency points by Fourier transforming this.

When the frequency positions are not evenly spaced, the Geltzel conversion method is used using the linear prediction coefficient α,
The power spectrum envelope value may be obtained. The power spectrum calculation means 101 may perform the following calculation. That is, the all-pole type spectrum model is expressed by the formula (1).
Can be represented by Here, α _i is the i-th order linear prediction coefficient, and σ is the average amplitude of the prediction error.

[0020]

[Equation 1]

The power spectrum at frequency ω is | H
(Ω) | ² . When the linear prediction parameter is an LSP parameter, θ _i is i-th order (i = 1, ..., P)
Assuming that the LSP parameter is, the power spectrum can be obtained by the equation (2).

[0022]

[Equation 2]

As described above, in this embodiment, the square root power spectrum envelope is obtained by the operation for obtaining the power spectrum, but only the envelope value of sparse frequency values is obtained. Therefore, the interpolating means 102 interpolates the square root power spectrum for the frequency position not obtained by the power spectrum calculating means 101 by the square root power spectra before and after the frequency root. In this embodiment, a place where the change in the square root power spectrum envelope is small is interpolated, and a place where there is a large change is linearly obtained by the spectrum envelope calculation means 103 using the linear prediction parameter obtained by the prediction parameter reproducing means 56. Representing the spectral envelope with a linear prediction parameter gives a slowly changing envelope characteristic. However, the number of peaks that is half the order of the poles occurs, and the approximation error becomes large near the peaks by approximation using only interpolation. Therefore, from the already calculated change of the spectral envelope value, the region of gradual change is interpolated, and the region of abrupt change is recalculated accurately. As a result, the approximation error can be kept small with a small amount of calculation. An actual example of this processing is shown in FIG. FIG. A is a conventional method for calculating the envelope value at all frequency points. B shows the case where the envelope value (x point) is calculated only at the even-numbered frequency positions, and C shows the case where the interpolation value (white circle) and the direct envelope value (black circle) are calculated. Here, the uncalculated portions of the portions that have relatively abrupt changes such as the calculated values a ₁ , a ₂ and a ₃ are actually calculated using the linear prediction parameter to obtain the value b ₂ , but the calculated value a ₂ , a ₃ ,
a _4, if relatively unchanged as a ₅ is gentle is interpolated using the calculated values before and after the uncalculated location obtaining an interpolated value _{c 1, c 2, c 3} . It should be noted that it is practical that the mottled calculation in the power spectrum calculation means 101 is performed at least every 5 points, or every 10 to 20 points.

Various criteria can be considered for the selection of the actual calculation or the interpolation. For example, when the spectrum envelope is modeled by the all-pole type, the peak of the spectrum envelope is sharper than that of the valley, and the change is large in many cases. By using this property, the amount of calculation of the spectrum envelope can be reduced without deteriorating the performance by finely calculating the upward convex region and interpolating in the downward convex region. If the spectral envelope value | H (Ω) | ² at the frequency point Ω is compared with the spectral envelope value at the position deviated by λ point, and if J given by equation (3) is positive, it can be regarded as convex. .

J = 2 | H (Ω) | ² − | H (Ω−λ) | ² − | H (Ω + λ) | ² (3) Further, this J and | H (ω) | ² are compared and ω The calculation of | H (ω) | ² from = Ω−λ to ω = Ω + λ may be performed according to the following rules.・ If J> 0.5 | H (Ω) | ² then actual calculation ・ 0.5 | H (Ω) | ² >J> 0.1 | H (Ω) | ² | H (Ω-λ) | ² and | H (Ω + λ) | ² is actually calculated only near the larger value. Others are interpolated. ・ If 0.1 | H (Ω) | ² > J, interpolated. Generally, the energy of the low frequency component is large and Since the fluctuations are large in many cases, it is also effective to calculate with a fine interval in the low frequency region and with a coarse interval in the high frequency region. For example, if the frequency is 2 kHz or less, the frequency is actually calculated for each frequency, but if the frequency is 2 kHz or more, appropriately sparse frequency points are calculated, and interpolation is performed between them. The calculated frequency interval may increase as the frequency increases.

A quadratic interpolation or a simple linear interpolation is sufficient as the interpolation method. The square root spectrum envelope value obtained from the power spectrum calculation means 101, the square root spectrum envelope value obtained from the interpolation means 102, and the square root spectrum envelope value obtained from the spectrum envelope calculation means 103 are combined and multiplied by the multiplier 82. Bowl 61
More frequency domain residual signal is multiplied and inversely normalized. Subsequent processing is the same as in FIG.

In the above description, the square root power spectrum envelope values at predetermined sparse frequency positions are obtained, and the uncalculated frequency points are interpolated or directly calculated.
All may be obtained by interpolation. Further, in the above, the square root power spectrum envelope on the encoding side normalizes the frequency domain signal and then quantizes it.However, after normalizing the frequency domain signal on the power spectrum envelope, the decoding of the quantized code, that is, FIG. The present invention can be applied to the decoding shown in FIG. In this case, the power spectrum envelope values at sparse frequency positions are obtained from the linear prediction parameters reproduced by the power spectrum calculation means 101, and for these,
As shown in FIG. 4, the calculation of the square root and the calculation of the reciprocal number are required, but the calculation amount can be reduced as compared with the case where such calculation is performed for all frequency points. Further, the transformation from the frequency domain to the time domain is not limited to the inverse modified discrete cosine transform, and other methods such as an inverse discrete cosine transform and an inverse discrete Fourier transform (inverse fast Fourier transform) may be used.

[0028]

As described above, according to the present invention, the square root of the power spectrum envelope is not obtained for all frequency points from the inversely quantized linear prediction parameter, but a part thereof is omitted and the omitted frequency is omitted. Since the points are obtained by interpolation, the number of calculations of the equations (1) and (2) is small, and since the equations (1) and (2) include division, the calculation amount is considerably reduced as a whole. In addition, for example, when Fourier transform is performed and the reciprocal thereof is obtained to obtain the power spectrum envelope value, the amount of subtraction of the Fourier transform is reduced, and the number of times the reciprocal is taken thereafter is also reduced, so the amount of calculation is considerably reduced. .

Therefore, according to the present invention, the spectral envelope for each frequency component is calculated by operating on sparse frequency points of, for example, 10 to 20 points with almost no increase in quantization distortion in conversion coding of speech or musical sound. Since the number of value calculations and divisions can be significantly reduced from about 1/5 to about 1/10, the calculation processing of the decoder can be reduced. In particular, in a signal processor, the division generally requires 20 to 30 times as many calculation steps as the multiplication, so that the calculation amount is greatly reduced. Further, according to the decoding for the code quantized after being normalized by the square root power spectrum envelope as in the embodiment, the square root calculation is also eliminated, and the number of calculation steps is remarkably reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a decoder to which an embodiment of the present invention is applied.

FIG. 2A is a diagram showing a square root power spectrum outline in which all frequency points in the conventional method are obtained from the reproduction linear prediction parameters, and B is a reproduction linear prediction parameter for only even points as sparse frequency points in the present invention. The figure which shows the calculated power spectrum outline, C is the figure which shows the square root power spectrum envelope outline example which interpolated the uncalculated frequency point with respect to the power spectrum outline calculated | required by B by this invention, and was actually calculated. Is.

FIG. 3 is a block diagram showing a decoder as an encoder to which a conventional transform coding / decoding method is applied.

FIG. 4 is a diagram showing how a power spectrum envelope value is obtained from a linear prediction coefficient by Fourier transform.

FIG. 5 is a block diagram showing an example of an encoder that normalizes and then quantizes a proposed square root power spectrum envelope.

6 is a block diagram showing an example of a decoder for the encoder of FIG.

Claims (6)

[Claims] 1. A residual signal is obtained by dequantizing a first index in an input code, a square root of a power spectrum envelope is obtained from a second index in the input code, and the residual signal is denormalized to obtain a frequency. An acoustic signal conversion / decoding method for obtaining a domain signal and transforming the frequency domain signal into a time domain signal to obtain an acoustic signal, wherein the linear prediction coefficient is dequantized from the second index to obtain a linear prediction parameter. A second step of obtaining a square root value of a power spectrum envelope at predetermined sparse frequency positions in the linear prediction parameter, and a square root value of the power spectrum envelope obtained in the second step. And a third step of interpolating the square root value of the power spectrum envelope at the frequency position that was not obtained in step 3, and Method of. 2. The method according to claim 1, further comprising a fourth step of directly obtaining a square root value of a power spectrum envelope of a part of frequency positions not obtained in the second step from the linear prediction parameter. The acoustic signal conversion decoding method described. 3. The acoustic signal according to claim 2, wherein the interpolation in the third stage is performed for a high frequency region, and the direct calculation in the fourth stage is performed for a low frequency region. Conversion decoding method. 4. The method further comprises a fifth step of judging the magnitude of a change rate of the square root of the power spectrum envelope obtained in the second step, and the third step at the frequency position determined to have a small change rate in the fifth step. 3. The acoustic signal conversion decoding method according to claim 2, wherein the interpolation is performed and the frequency position determined to have a large change rate is directly obtained in the fourth step. 5. The acoustic signal conversion decoding according to claim 1, wherein in the second step, a square root of a power spectrum envelope is obtained as a result of performing an operation for obtaining a power spectrum envelope from the linear prediction parameter. Method. 6. The method according to claim 1, wherein in the second step, an operation for obtaining a power spectrum envelope is performed from the linear prediction parameter, and a square root of the operation result is obtained to obtain a square root of the power spectrum envelope. 4. The acoustic signal conversion decoding method described in 4.