JP3134363B2 - Quantization method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a quantization method applied to a signal processing system for transmitting a speech signal after encoding it with high efficiency, and more particularly to a quantization method for additional information of a speech signal.

[0002]

2. Description of the Related Art In a high-efficiency coding of an audio signal (audio signal), an input audio signal is divided into a plurality of channels along a time axis or a frequency axis, and bits for adaptively allocating the number of bits for each channel. There is a coding technique based on allocation (bit allocation). For example, coding techniques based on bit allocation for audio signals and the like include:
Band division coding (sub-band coding: SBC), which divides an audio signal on the time axis into a plurality of frequency bands and encodes the signal, and converts (orthogonal transform) a signal on the time axis into a signal on the frequency axis. So-called adaptive conversion coding (AT) in which the frequency band is divided into a plurality of frequency bands and adaptively encoded for each band.
C) or a combination of sub-band coding and so-called adaptive predictive coding (APC), dividing the signal on the time axis into bands, converting each band signal to baseband (low band), and then performing multi-order linear prediction. There are coding techniques such as so-called adaptive bit allocation (APC-AB) for performing analysis and predictive coding.

[0003] Among these high-efficiency codings, for example, in adaptive transform coding, an audio signal on the time axis is transformed by an orthogonal transform such as a fast Fourier transform (FFT) or a discrete cosine transform (DCT). To an axis (frequency axis) orthogonal to the time axis, and then divides the band into a plurality of bands, and quantizes (requantizes) the FFT coefficients, DCT coefficients, and the like in each of the divided bands by adaptive bit allocation. Conversion)
are doing. As an example of requantization in adaptive transform coding of the fast Fourier transform, as shown in FIG. 9, for example, an FFT amplitude value Am or the like after a signal is subjected to fast Fourier transform is divided into blocks (blocks B1, B2 ‥‥). Then, there is a method of calculating additional information necessary for requantizing each of these blocks, requantizing each block using the additional information, and quantizing the additional information itself.

[0004]

In high-efficiency coding in which the orthogonal transform is performed on an axis orthogonal to the time axis, a masking threshold is generally obtained from the power of the spectrum within the bark, and the masking threshold is calculated. In order to suppress quantization noise below a threshold level, dynamic bit allocation has been performed on the frequency axis. Here, the width of one bark is determined by the characteristics of human hearing (the ability to recognize humans), and by performing the above-described processing for each bark, an auditory masking effect (simultaneous masking) is used. High efficiency coding is performed.

In this case, an RMS value (floating coefficient) and a masking threshold value are added as additional information of the bark spectrum for each bark, and this additional information is also quantized. This additional information is quantized by a fixed uniform bit allocation independently for each sample in the decibel region (on the log axis) or normalized by a peak in a frame or a frame gain and then quantized. However, parameters such as the RMS value and the masking threshold value change significantly at different frequencies, so quantizing with fixed uniform bit allocation is very inefficient.

[0006] Therefore, in order to be able to compress audio data at a higher compression rate, it is required to efficiently compress additional information.

An object of the present invention is to provide a quantization method capable of compressing additional information at a high compression rate.

[0008]

SUMMARY OF THE INVENTION The present invention provides a quantization method for quantizing additional information of a highly efficient coded audio signal when the audio signal is highly efficiently encoded by orthogonal transformation. A linear approximation of the parameter is obtained, and a deviation from this linear approximation is quantized.

[0009]

In this manner, the additional information is also efficiently compressed and quantized, and the coding is performed with higher efficiency, so that the transmission rate can be further reduced.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to FIGS.

In this embodiment, a transmitting side (encoder side) for performing high-efficiency encoding of an audio signal is configured as shown in FIG. 1, and a receiving side (decoder side) of the high-efficiency encoded audio signal is illustrated in FIG. It is configured as shown in FIG.

First, a description will be given of a configuration of an encoder for encoding a voice signal with high efficiency. In FIG. 1, reference numeral 1 denotes an input terminal of a voice signal, and a digital voice signal (digital audio signal) obtained at the input terminal 1 is shown. Is supplied to the pitch prediction circuit 2 and the data on which the pitch prediction on the time axis has been performed is supplied to the windowing / orthogonal transformation circuit 3, where appropriate windowing is performed and orthogonal transformation is performed. For windowing at this time, the sampling frequency f _S of the input signal is set to, for example, 4
If the frequency is 8 kHz, for example, as shown in FIG.
Windowing is performed with the sample as one unit. In this case, the overlapping range is set to 1/16, and 64 samples are overlapped.

The orthogonally transformed data is supplied to a spectrum intensity calculation circuit 4 to calculate the intensity (power) of each spectrum. In this example, since the fast Fourier transform (FFT) processing is performed, the (real part) ² +
(Imaginary part) Calculate the intensity with ² . The intensity data of each spectrum at this time is supplied to the Bark integration circuit 5, where it is integrated for each critical band, and the intensity of the Bark spectrum is calculated. At this time, integration by the following equation is performed.

[0014]

(Equation 1)

Then, based on the intensity distribution of the bark spectrum, a masking threshold and RMS value calculation circuit 6 calculates a masking threshold value for each bark band. In this case, the absolute value is determined in consideration of the absolute threshold determined from the minimum audible curve of hearing. The value of the masking threshold found here is divided by the bandwidth of each bark, then squared,
Set the effective value for each sample. The RMS value (floating coefficient) is also obtained by the masking threshold and RMS value calculation circuit 6.

Next, the masking threshold and R
Threshold value and RMS obtained by MS value calculation circuit 6
The value is supplied to a parameter quantizer 7 where the threshold value and the RMS value are quantized. At this time, the RMS value is obtained by dividing the intensity of each bark by the number of samples in the bark, taking a square root, and quantizing the RMS value as an effective value per sample.

In this embodiment, a linear approximation of the threshold value and the RMS value is obtained, and a deviation from each linear approximation is quantized. Hereinafter, this quantization method will be described. Here, the quantization of the RMS value of the bark spectrum will be described as an example.

First, the bark spectrum will be described. No. 1 to No. There are 24 lines up to 24 (see Table 1 described later), and the values of the bark spectra of these 24 lines are represented by B ₁ , B ₂ ‥‥ as shown in FIG.
B ₂₄ , these 24 bark spectra B _{1 to} B
₂₄ is represented by the following equation.

[0019]

(Equation 2)

In the equation (2), i is a value corresponding to each bark spectrum of 1 to 24, k is a value of a sample point (for example, a value of 1,2,1024), and Y _(k) is a coefficient ( F
FT coefficient, DCT coefficient, etc.), u _(i) is the upper limit of the i-th bark spectrum, and l _(i) is the lower limit of the i-th bark spectrum.

Here, if the rsm value (floating coefficient) per sample for each bark is defined in a decibel (dB) region, it is expressed by T _(i) in the following equation.

[0022]

(Equation 3)

The rsm value T _{(i) represented} by the equation (3 ₎
Is linearly approximated by the least squares method, and a linear approximation indicated by a two-dot chain line in FIG. 4 is obtained. Next, general straight line approximation by the least squares method will be described. First, the following equation is calculated.

[0024]

(Equation 4)

The linear approximation is performed by obtaining the values of a and b shown in the equation (4). That is, the following [Equation 5]
Equation and Equation 6 may be solved.

[0026]

(Equation 5)

[0027]

(Equation 6)

Solving both equations gives the following.

[0029]

(Equation 7)

[0030]

(Equation 8)

Then, the following equation is obtained from the equation (8).

[0032]

(Equation 9)

By substituting the equation (9) into the equation (7), the following equations (10) and (11) are obtained.

[0034]

(Equation 10)

[0035]

[Equation 11]

[Equation 10] and [Equation 11]
In the equation, as conditions for obtaining a linear approximation of the Bark spectrum of this example, x _i → i (i is 1 to 24), y _i → T _(i), N =
By substituting 24, the following equations [Equation 12] and [Equation 13] are obtained.

[0037]

(Equation 12)

[0038]

(Equation 13)

From the values of a and b obtained by the equations (12) and (13), the i-th value Tl _(i) = ai + b on the straight line approximated by a straight line is defined. At this time, a corresponds to the inclination of the straight line, and b corresponds to the offset value. This Tl _{(i) is} also a value expressed in decibels (dB). This linear approximation operation can be executed by a Lloyd Max Quantizer of about 8 bits. Alternatively, quantization can be simply performed by a uniform quantizer.
Then, if the quantized values of a and b are a 'and b',
The value Tl ' _(i) of the quantized straight line is represented by the following equation (Equation 14).

[0040]

[Equation 14]

The value Tn quantized in this embodiment is
_(i) is shown by the following equation (Equation 15).

[0042]

(Equation 15)

Here, T _(i) is a true value, and Tn _{(i) in} equation ₍ 15 ₎ indicates that T _(i) is normalized by Tl ′ _(i). I have. Where Tn
_{Since (i)} has a narrower dynamic range than T _(i) , it can be seen that quantization can be performed with a smaller number of bits.

Then, based on the threshold value and the RMS value ratio quantized by the parameter quantizer 7 in this manner, the bit allocation calculation circuit 8 allocates the number of bits to be assigned to one sample in each bark (ie, [Equation 15]). Tn _{(i) in} the equation). Next, a method of calculating the bit allocation in the bit allocation calculation circuit 8 will be described.

As the calculation of the bit allocation suitable for this example, there are a case where a fixed and uniform bit allocation is calculated, a case where a fixed non-uniform bit allocation is calculated, and a case where a dynamic bit allocation is calculated. .

First, the case of calculating a fixed and uniform bit allocation will be described. In this case, the distribution of Tn _(i) is checked for each bark i, and the power distribution (distribution for a relatively long time) is obtained. Is assigned in proportion to. At this time, for example, as shown in FIGS. 6A and 6B (A and B indicate continuous orthogonal transform frame portions),
For each frame (orthogonal transformation frame), the slope and T
n _(k) changes. Here, assuming that the k-th value of the j-th frame is Tn _(k) j, this value Tn _(k) j is represented by the following equation [Equation 1]
6].

[0047]

(Equation 16)

In the equation (16), k is from 1 to 24.
And the value of this [Equation 16] is the value of T in J frames.
σ (standard deviation ₎ of n _(k) . However, this is a standard deviation in dB value.

In this fixed bit allocation method, one bit is given to the maximum standard deviation σ, and [new standard deviation σ _(NEW) ] ← [original standard deviation σ _(OLD) −6]. Next, the values of σ are arranged in a line and the largest is set to 1
Bits are given, and the value obtained by subtracting 6 dB is arranged in a row as a new value of σ. Then, the total bits (for example, 4 × 25
(= 100 bits) until a fixed bit allocation is determined.

As another fixed bit allocation, there is a method of making the dB value per step constant. That is, one bit is given to the largest σ, and [new standard deviation σ _(NEW) ] ← [original standard deviation σ _(OLD) / 2]. That is, what is given by 1 bit is divided by 2 and arranged in a line as a new value. Then, the following is repeated in the same manner as the above-described fixed bit allocation to determine a fixed bit allocation.

As yet another fixed bit allocation, there is a method of using a gradient a of a straight line Tl _(i) = ai + b. In this case, assuming that the average of a in J frames (J is a sufficiently large value) is taken as a ″, this a ″ indicates the average T _(i) slope in the dB region. That is, i
Represents an increase in the dB value every time the value of the increases by 1. Therefore,
For Tl ₍₁₎ and Tl ₍₂₄₎ , nint ｛(a ″ × 23)
It is necessary to provide a / 6 bit drop. nint indicates the nearest integer. This drop is d bits. However,
Let d <0. By doing so, the following equation (Equation 17) is obtained.
The total number of increased bits C may be determined so as to satisfy the following.

[0052]

[Equation 17]

The increased bit number C obtained by the equation (17)
FIG. 7 shows the relationship between the data and the difference d bits.

Next, the case where the dynamic bit allocation is calculated will be described. In this case, the following formula [Equation (2)] is used for each frame without using the average value a ″ of the gradient a in the J frames described above. 18].

[0055]

(Equation 18)

The total bits in the expression (18) are the total bits of side information (additional information). Then, the increased bit number C is determined so as to satisfy the expression (18).

When the above-described bit allocations are evaluated, the fixed bit allocation based on [Equation 16] becomes flat because the variance of the difference from the straight line is used as the reference for the bit allocation. In the bit allocation based on the expression or the expression (18), since the rough variance of the straight line itself is used as a reference for the bit allocation, more bit allocation is performed in the low frequency band.

Here, returning to the description of the entire configuration of the encoder with reference to FIG. 1 again, the threshold value and the RMS value quantized by the parameter quantizer 7 in this way and the bit allocation calculation circuit 8 determine the threshold value and the RMS value. The value of the number of bits allocated to one sample in each bark is supplied to an adaptive quantizer 10 described later.

Then, the data subjected to the orthogonal transform by the windowing / orthogonal transform circuit 3 is supplied to a fast Fourier transform (FFT) circuit 9 to perform the decimation (averaging, smoothing) within the critical band. Then, the data decimated by the fast Fourier transform circuit 9 is applied to the adaptive quantizer 1.
0 to perform quantization. Here, in this example, the adaptive quantizer 10 quantizes a representative value in each critical band.

Then, the representative value in each of the quantized critical bands is transmitted from the representative value output terminal 11.
The threshold value and the RMS value quantized by the parameter quantizer 7 as the additional information are transmitted from the additional information output terminal 12.

Next, data processed by the encoder having such a configuration will be described. First, a list of each bark (that is, each critical band: critical band) is shown in Table 1.

[0062]

[Table 1]

Here, for example, the sampling frequency f _S =
Assuming that the FFT of 1024 samples is performed at 32 kHz, the frequency range f of 0 to 16 kHz is 512
There is a spectrum of points. At this time, the spectrum spacing is 31.25 Hz (= 16000/512).
Becomes Here, in the case of an audio signal, the main components are mostly concentrated to 5 kHz or less, and especially 2 kHz to 3 kHz.
Energy is concentrated at kHz.

Now, there is a critical bar at 1 kHz.
If you think about Bund (Bark), from the above [Table 1]
No. There are 1 kHz in 9 critical bands.
This No. 9 critical bands from 920Hz
There is a range up to 1080 Hz, and the spectrum
There are six. For example, as shown in FIG.
Vector x₁, X_Two, X_Three, X_Four, X_FiveSuppose there was
You. At this time, in the encoder of this example, as shown in FIG.
Thus, the five smoothed spectra y₁, Y _Two,
y_Three, Y_Four, Y_FiveAnd And in this example,
Y₁= Y_Two= Y_Three, Y_Four= Y_FiveHas become
And the spectrum y₁And the spectrum y_FourOnly the quantized value of
Is transmitted as a representative value.

Here, from the spectrum x to the spectrum y
An example of the conversion process of
Within the sliding band (x in FIG. 8A)₁~ X_ThreeBand and x_Four
~ X _FiveEnergy in the band) is affected by the conversion
Since it is necessary to avoid the above, Equation 19 and
The processing of Expression 20 is performed.

[0066]

[Equation 19]

[0067]

(Equation 20)

From equations (19) and (20),
The spectrum y ₁ and the spectrum y ₄ are shown by the following [Equation 21] and [Equation 22].

[0069]

(Equation 21)

[0070]

(Equation 22)

Equations (21) and (22) show
Spectrum y₁And the spectrum y _FourAnd each sample
The effective value per hit. Then, this representative value y₁, Y_Four
The relationship between the rms value of the critical band and
Is done.

[0072]

(Equation 23)

That is, if one critical band is divided into n subbands and smoothed, the nth representative value is calculated from the RMS value of the entire critical band and the (n-1) representative values. Is found. Then, the representative value is quantized with the number of bits per sample determined based on the quantized threshold value and the RMS value, thereby determining data to be transmitted.

Here, an example of smoothing is shown in the following [Table 2].

[0075]

[Table 2]

The bark No. shown in Table 2 was used. Is the bark No. in [Table 1]. It corresponds to. In this [Table 2], two examples of numbers to be concluded by smoothing are shown as smoothing example 1 and smoothing example 2, and for example, N shown in FIG.
o. In the case of 9 critical bands, it can be seen from smoothing example 1 that two spectra are bundled and three spectra are bundled. In addition, No. In the case of the critical band 13 or later, smoothing example 1 and smoothing example 2 can be selected. N from 9 critical bands
o. No. 12 is included in the smoothing example 1 up to the critical band. For critical bands 13 and thereafter, it is preferable to classify them in Smoothing Example 2.

As can be seen from Table 2, although it is originally necessary to transmit the quantized values of 512 spectra obtained by summing the number of spectra of each critical band in all bands, only the smoothing example 1 is used. In this case, only the quantized values of 62 spectra need to be transmitted. Even when the smoothing example 1 and the smoothing example 2 are combined as described above, 1
It is only necessary to transmit the quantized values of the 04 spectra.
Therefore, the data amount of the transmission signal can be significantly reduced, and high-efficiency encoding with a higher compression ratio is performed.

In this embodiment, the quantization of the additional information in the parameter quantizer 7 is performed by obtaining a linear approximation of each parameter and then quantizing the deviation from the linear approximation. Efficient quantization using parameter correlation is performed. Therefore, the additional information is also efficiently compressed and quantized, and more efficient encoding is performed, so that the transmission rate can be further reduced. In this case, the energy of the total quantization noise (that is, the quantization noise at the time of quantization of the representative value) can be reduced by setting the non-uniform bit allocation according to the RMS and the variance of the parameter. In addition, since the dynamic bit allocation is performed according to the parameter variance, good quantization that follows a change in the input signal is performed, and the conversion efficiency is improved from this point as well.

In this embodiment, the compression processing is performed after the pitch prediction of the input signal is performed, so that the high-efficiency coding is performed more effectively.

Next, a decoder for receiving the data smoothed and quantized as described above will be described with reference to FIG. 2. In FIG. 21, reference numeral 21 denotes a representative value quantized value in each critical band. Indicates the representative value input terminal
Reference numeral 2 denotes an additional information input terminal to which additional information (threshold value and quantized value of RMS value) of this value is transmitted. Then, the data obtained at both input terminals 21 and 22 are supplied to an adaptive inverse quantizer 23 to obtain a representative value in each critical band, and this representative value is subjected to an interpolation process by a coefficient interpolation circuit 24. At this time, it is necessary to keep the energy in each critical band unchanged. For example, the representative value is repeatedly used as it is for interpolation. Then, the interpolated data is supplied to an inverse transform / window superimposing circuit 25, and the frequency axis is inversely transformed to a time axis and the windowed data is superimposed. Then, this inverse conversion / windowing superposition circuit 2
The data processed in (5) is supplied to the inverse pitch prediction circuit 26 to restore the original digital audio signal, and this digital audio signal is supplied to the audio signal output terminal 27.

The digital audio signal decoded in this way has the same energy in each critical band as the original audio signal, so that the auditory sound quality degradation when reproducing this audio signal is minimized. In fact, there is practically no deterioration in sound quality because there is a reduction in the ability to identify frequency components of hearing. That is, in human hearing, if the energy within one bark does not change, it is difficult to identify the position of the spectrum within that bark.
Even if the voice transmitted after being subjected to the high-efficiency coding described above is reproduced, the sound quality is not substantially deteriorated.

In the above embodiment, the masking threshold value is transmitted as the additional information. However, if the threshold value is formed from the quantized RMS value, the decoder side can also obtain the threshold value from this RMS value. Is possible, the threshold value does not have to be transmitted. In this case, the transmission rate of the additional information is reduced, and the transmission rate can be further reduced.

In the above embodiment, the compression process is performed after the pitch of the input signal is predicted.
The compression processing may be performed without performing pitch prediction. However, when the pitch prediction is performed, high-efficiency coding is performed more effectively.

The spectrum integrated by the bark integration circuit 5 on the encoder side is subjected to audibility correction such as loudness conversion, and then the masking threshold and RMS are corrected.
It may be supplied to the value calculation circuit 6.

In the above-described embodiment, the orthogonal transform is performed by the FFT. However, the present invention can be applied to a high-efficiency encoding that performs other transform processing. For example, the present invention can be applied to high-efficiency coding by DCT (Discrete Cosine Transform) or MDCT (Modified DCT). In this case, for example, when DCT is applied, the calculation in the spectrum intensity calculation circuit 4 of the encoder is performed only with (real part) ² (in the case of DCT, there is no imaginary part). Also, in the case of DCT, the integral processing of the spectrum intensity is performed by the following equation instead of the equation (1).

[0086]

(Equation 24)

Further, in the above-described embodiment, no description has been given of the transmission system of data which has been efficiently coded by the encoder. However, various transmission systems such as a wired system and a wireless system can be applied. The present invention can also be applied to a case where after the converted data is recorded on various recording media, a reproduced signal from the recording medium is restored by a decoder.
In any case, since the bit rate is greatly reduced in this example, the transmission efficiency (recording efficiency) is good.

[0088]

According to the present invention, when quantizing the additional information, a linear approximation of each parameter is obtained, and the deviation from this linear approximation is quantized. Therefore, the additional information is efficiently compressed. Quantization is performed, and more efficient encoding is performed, so that the transmission rate can be further reduced.

[Brief description of the drawings]

FIG. 1 is a configuration diagram illustrating an encoder according to an embodiment of the present invention.

FIG. 2 is a configuration diagram illustrating a decoder according to an embodiment of the present invention.

FIG. 3 is an explanatory diagram showing a windowing state according to one embodiment.

FIG. 4 is an explanatory diagram showing an example of linear approximation.

FIG. 5 is a waveform chart for general explanation of linear approximation.

FIG. 6 is a waveform chart for explaining linear approximation according to one embodiment;

FIG. 7 is an explanatory diagram of bit allocation according to one embodiment.

FIG. 8 is an explanatory diagram showing an example of a spectrum according to one embodiment.

FIG. 9 is an explanatory diagram showing blocks of adaptive transform coding.

[Description of Signs] 1 audio signal input terminal 2 pitch prediction circuit 3 windowing / orthogonal conversion circuit 4 spectrum intensity calculation circuit 5 bark integration circuit 6 masking threshold and RMS value calculation circuit 7 parameter quantizer 8 bit allocation calculation circuit 9 Fast Fourier transform circuit 10 Adaptive quantizer 11 Representative value output terminal 12 Additional information output terminal 21 Representative value input terminal 22 Additional information input terminal 23 Adaptive inverse quantizer 24 Coefficient interpolation circuit 25 Inverse transformation / windowing superposition circuit 26 Pitch Inverse prediction circuit 27 Audio signal output terminal

Claims (1)

(57) [Claims] 1. A quantization method for quantizing additional information of a highly efficient encoded audio signal when the audio signal is highly efficiently encoded by orthogonal transform, wherein a linear approximation of the parameter of the additional information is obtained. A quantization method wherein a deviation from the linear approximation is quantized.