JPH06291679A - Threshold value control quantization determining method for audio signal - Google Patents

Abstract

PURPOSE:To obtain the threshold value control quantization determining method for the audio signal which performs bit allocation at low cost while holding the quantity of a recomposed signal excellent. CONSTITUTION:A filter bank 101 generates a subband sample from an input audio signal, and a maximum value is found from the subband sample, and quantized by a 1st quantizing means 103. In a process wherein the quantized maximum value is normalized and quantized by a 2nd quantizing means 108, the mean energy of the subband is calculated 110, a division coefficient is found 111 by specific arithmetic on the basis of the mean energy, and 2nd quantization is performed according to the division coefficient.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION This invention relates to efficient information coding of digital audio signals for transmission or digital storage media.

[0002]

2. Description of the Related Art The compression algorithms developed for high-fidelity audio in recent years are mainly frequency domain coders. These coders decompose the input digital audio signal into the frequency domain in order to extract spectral characteristics and remove redundant and improper ones. In most of these designs, the encoder
It can be even more complicated, but the decoder must be simple. The bit allocation plan of the encoder is
It plays a crucial role and determines the coder's complexity, the extent of redundancy elimination, and the quality of the reconstructed output.

Subband coding (SBC), which utilizes feedforward quantization for psychoacoustic modeling of audio signal coding, is an ISO standard.
/ WG 11 / MPG (Motion Picture Experts Group) forms the core method of the audio coding standard. The bit allocation procedure used in the subband coding algorithm is based on the masking threshold of the human ear, and the reconstructed output is transparent at bit rates above 128 kbit / s / channel. It turned out that sex could be obtained.

FIG. 6 shows a block diagram of a subband coder utilizing a perceptual bit allocation procedure. First, the filter bank 101 filters the input audio signal x into subbands to obtain subband samples S _k . From these samples, the maximum value determination means 102 determines the maximum value Smax at each prescribed time interval. Next 6
When the maximum value Smax obtained by the bit uniform quantizer (first quantizer) 103 is quantized, a quantized Smax is generated. In block 104, the quantized Smax is inverted and sampled with S _k by multiplier 1
When multiplied by 05, this normalization is performed before it is input to the second quantizer 108, and S _k for transmission is obtained. The second quantization step used for each subband is
It depends on the output of block 107 which calculates the bit allocation. Here, the input audio signal x is sent by another route to the block 106, and first, the block 106a
Is subjected to a fast Fourier transform (FFT) and is decomposed into spectral components at each time interval or frame. A masking threshold is then calculated by block 106b based on the simultaneous masking of the tone-to-noise and noise-to-noise. As a result, mask values are obtained for each subband and these mask values are used in the block 107 bit allocation procedure.

FIG. 7 shows a flow chart outlining the processing of blocks 106 and 107. First, step S71 is an FFT process and corresponds to block 106a. Step S72 corresponds to block 106b where the mask value is obtained. Steps S73, 74, 7
5, 76, 77, 78, and 79 indicate the processing steps within block 107. That is, the quantization noise is calculated by using the mask value and the quantized Smax (step S73), and the noise-to-mask ratio is calculated by the quantization noise (step S74). The resulting noise-to-mask ratio is used in an iterative method to obtain the number of bits allocated to each subband below. The steps in the iterative method are as follows.

[1] Maximum noise to mask ratio (NMR)
Then, the frequency band i is calculated (step S75).

[2] The bit assigned to the frequency band Li is incremented (step S76).

[3] Recalculate NMRi (step S
77).

[4] Remaining bit B where B = BM
Is calculated (step S78).

[5] When B is M or more, the above process [1]
To [5] are repeated, otherwise allocation is finished (step S79).

Also, in May 1985, Proceedings of ICASSP'82, Paris, 203-206.
TA Ramstad, described on page
By "Subband Corder With A Simple
Adaptive Bit Allocation Algorithm Possible Candy Date Digital Mobile Telephony (Subband coder with a simple
adaptive bit allocation algorithm-a possible candi
A paper entitled "date for digital mobile telephony)" describes a simple bit allocation method based on the relative standard deviation of subband signals. The bit allocation of this method is based on the principle of simple majority voting, the band with the highest standard deviation receives the first bit, and then the standard deviation decreases by a linear coefficient. In addition, Wai. Another variation of the bit allocation procedure is found in the patent specification entitled Y. Matsui's method of encoding digital audio signals. In FIG.
Its block diagram is shown. 201 is a level calculation block in which the level of the subband signal is determined. Two
02 is a logarithmic value calculation block, and 203 is a scale
This is a factor calculation block. Block 204 holds the weighted values of the logarithmic table, and block 205 contains the weighting table for the subbands. Block 206 based on the linear weighted ratio of the logarithmic levels.
At, a bit allocation calculation is performed.

[0012]

However, the above-described perceptual bit allocation procedure has been tested at 12
Although transparency was obtained at bit rates of 8 kilobits / second and above, its complexity suffers. The FFT and mask threshold calculations are intensive and require out-of-order processing power, which poses the problem of limited implementation effectiveness at low cost.

On the other hand, the simple bit allocation method is insufficient in that bits are not allocated in consideration of the human auditory system. That is, firstly, the allocation is based on standard deviation, not on energy in the band, which is directly related to the perceived loudness. Second,
Since the same weighting is applied to the high-frequency subbands as to the low-frequency subbands, there is a problem in that even bits are distributed to the high-frequency subbands whose ear sensitivity is low.

The above can be seen, for example, from the results of example frames of audio signals for perceptual bit allocation and simple bit allocation shown in Table 1. The required bit allocation, calculated as the minimum bit allocation required to ensure that the quantization noise is below the mask, is also listed side by side for comparison. These bit allocations are 192 kilobits /
It is derived from 32 equal band sub-band code decoders operating at a bit rate of seconds / channel. As is clear from the bit allocation results, the number of bits allocated by the simple bit allocation procedure is less than that required for bands 0, 1 and 2. As can be seen, there is audible noise in these three subbands. The probability of hearing noise in this frame of the audio signal is 3/32. On the other hand, in the case of perceptual results, it can be seen that the allocated bits always exceed the required number.

[0015]

[Table 1]

In consideration of the above problems of the prior art, the present invention is a threshold control quantization decision for an audio signal which can be bit-allocated at a low cost while maintaining a good quality of a reconstructed signal. It is intended to provide a method.

[0017]

According to the present invention of claim 1, a sub-bank sample is generated from an input audio signal by a filter bank, a maximum value is obtained from the sub-band sample, and the maximum value is first quantizing means. Quantized by
In the process of normalizing the quantized maximum value and quantizing it by the second quantizing means, the average energy of the subband samples is calculated, and the division coefficient is obtained by a predetermined calculation based on the average energy, It is a threshold control quantization determination method for an audio signal which is secondly quantized according to the division coefficient.

The invention of claim 2 provides a defined time interval in determining the quantization of a digital audio signal with spectral and temporal structure, each represented by a plurality of frames of spectral quantized components. The step of obtaining the peak value of the frame component in
By comparing the peak value with the hearing threshold in the human auditory system, the steps of removing the components in the frame, calculating the average energy of the frame within the time interval, and the predetermined relationship with the average energy A threshold control quantization determination method for an audio signal, comprising a step of changing a frame quantization step.

[0019]

The present invention calculates the average energy of subband samples, obtains the division coefficient by a predetermined calculation based on the average energy, and performs the second quantization in accordance with the division coefficient. Can be assigned,
The quality of the reconstructed signal can also be improved.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings showing its embodiments.

FIG. 1 is a block diagram for realizing a threshold value control quantization determination method according to a first embodiment of the present invention. Those having the same numbers as those in the conventional example in FIG. 6 indicate processing blocks having the same function. 109 and 110 are
A power estimation block for calculating the average energy of the subband samples is represented, and 111 is a block for performing bit allocation based on the average energy (power).

FIG. 2 is a flow chart of the algorithm in the block diagram of the above configuration. Power estimation block 10
9, 110 corresponds to step S1 and corresponds to the block 111.
Corresponds to steps S2-9. The processing steps in FIG. 2 are elaborate as follows.

[1] The energy σ _i is calculated for each subband i, for a block of samples S _k of the input subband, ie σ _i = ΣS _k . As a result, a decibel value is obtained (step S1).

[2] Assuming zero bit allocation,
The quantization noise N _{i in} dB is calculated from the maximum value Smax for each subband (step S2).

[3] It is checked whether the hearing threshold Q _i exceeds the noise N _i (step S3). This hearing threshold is static and does not require additional calculation time for its calculation.

[4] If it is exceeded in the processing step [3], the energy σ _{i of the} subband is set to a low dummy value (step S4). For example, 10
Since it is set to a value 0 dB lower, during the iterative process,
Bits are most likely to be assigned to that subband.

[5] Search for subband i having the maximum average energy (step S5).

[6] Sub-bit bit allocation is Li
Only increments (step S6).

[7] When different γ _n are used for different clusters of subbands, the average energy is reduced by γ _n (step S7). These values γ _n are optimized to obtain the best subjective sound quality. For example, γ of subbands 0 to 4 is given the value 6 and γ of subbands 5 to 31 is given the value 8.

[8] Calculate the number of bits used (step S8). That is, M = kLi, where k is the number of time samples in the subband and Li is the number of incremental bits for subband i.

[9] When the remaining number of bits becomes smaller than the next possible number of bits, the bit allocation is completed. Otherwise, steps S5, 6, 7, 8 and 9 above are repeated (step S9).

FIGS. 3 and 4 are flow charts of the second and third embodiments of the present invention taking the calculation time into consideration.
The most likely use of this bit allocation procedure is in cost-sensitive products, thus providing an iterative procedure alternative if the computational speed of the processor is limited.

FIG. 3 shows the bit allocation procedure when the available calculation time by accurate estimation is converted into the number of bits that can be iteratively allocated. The remaining bits are used as a "seed" bit allocation for the subband. These "seed" values can be assigned in weighted proportions, either directly proportional to the subband energy or to reflect the ear's sensitivity to different bands. If the calculation time cannot be estimated, a count is needed in the iterative loop to indicate whether the calculation time has expired. When finished, the remaining bits are allocated as in the "seed" bit allocation above.

In FIG. 3, each step S1, S2, S
3 and S4 are steps S1, S2,
The process is the same as S3 and S4. Step S10 is
The determination of the seed ratio is represented, step S11 shows the allocation by the weighted ratio of energy, and step S12 represents the steps S5, S6, S in FIG.
This corresponds to the iterative process of 7, S8, and S9. The following is the details of the process of step S10 performed in determining the seed ratio regarding the flow of information.

1.1 Calculate the time t at which the processor is used in the iterative loop.

2. Assuming there are processor restrictions,
Find an acceptable time limit T.

3. Assuming that there is a time constraint T, the number N of allowable loops is calculated.

4. If one bit is allocated for each loop, N is the number of available bits. N is a seed ratio related to the flow of information.

At the next step S11, the initial allocation Li of each sub-band is obtained by Li = w _i σ _i / Σw _i σ _i ^* N. Here, w _i is a weight given to subband i.

In FIG. 4, step S13 represents the processing of steps S1 to S4 in FIG. 2, and step S14 represents the processing of steps S5 to S8 in FIG. In step S15, it is checked whether the number of available bits has been used up. Step S16
Then, by counting the number of iteration loops, it is checked whether the calculation time has ended. If the calculation time has not ended, the process returns to step S14 and is repeated. When the calculation time is over, the remaining bits are allocated based on the ratio of the remaining energy (step S17). The processing here is similar to step S11 in FIG. 3 except that the average energy value σ _i used is modified non-linearly by an iterative process.

The reconstructed output results showed almost as good subjective sound quality as the perceived coder. Utilizing the probability of noise measurement, the γ value exhibits a zero probability for a sequence. It is also possible to deliver an output showing an average of 3.5% probability for sequences tested at 128 kbps. Table 1 shows the bit allocation of the present invention for the same frame of the audio signal described above. All energies are divided into subbands using only a single γ value. Subband 1
Improvements were observed in 2 and 2 and the probability of noise generation is now reduced to 1/32. FIG. 5 shows a bit allocation comparison for another frame of the audio signal. The results of the invention, labeled "threshold", indicate that it is better suited for perceptual allocation than simple bit allocation. A simple bit allocation indicates that subbands 0, 1 and 2 can be noisy, and the result of the threshold control bit allocation is to meet the minimum allocation required by masking theory. It means that. The present invention is 12
Bit rates of 8 kilobits / second and above have been found to be most effective.

As described above, according to the present invention, the simplicity of the simple bit allocation is maintained, and at the same time, in order to further remove the improper ones, the human auditory system is taken into consideration to correct the shortage. Is added. That is, first, the peak value of the subband sample is obtained at each prescribed time interval. Second, by applying a hearing threshold and comparing it to the peak value, psychoacoustically inappropriate subbands are removed. This often results in a narrow audio signal bandwidth. The hearing threshold is the lowest sound pressure level of a pure tone that the human ear can hear, which is a function of frequency. Third, the average energy of each subband is calculated and used as the basis for allocation. Finally, the bits are repeatedly allocated for each subband. If a bit increment is assigned to a subband, for each group of subbands,
A non-linear value is assigned as the division factor for the average energy. This is done in consideration of the different ear sensitivities at different frequencies and sound pressure levels.

Also, the number of bits to be assigned to each subband is calculated using the average energy and peak value in the band, the worst case value regarding the threshold value during silence, and the nonlinear reduction coefficient. This bit allocation technique does not affect the structure of the encoding algorithm, so the decoding algorithm can remain the same. To improve the quality of the reconstructed output, the non-linear reduction factor can be weighted for different clusters of subbands. Furthermore, if the calculation time is constrained, it is also possible to first calculate an initial value or "seed" value for each subband before iteration.

From the above, the strength of the present invention lies in its simplicity and means for obtaining a quality comparable to a coder using perceptual bit allocation. According to the ISO / MPEG audio subgroup rating, an encoder with the configuration of Figure 6 cannot be implemented in a single DSP due to the complexity of the algorithm. This has made cost an important issue, and is deterring the use of encoders in consumer products where very low bit rates are not required. The present invention is of great importance as a low cost alternative due to its simplicity and effectiveness in obtaining good sound quality.

In the above embodiment, the blocks for performing the respective processes were constructed by dedicated hardware, but instead of this, similar functions may be constructed by software using a computer.

[0046]

As is apparent from the above description, the present invention calculates the average energy of subband samples, obtains a division coefficient by a predetermined calculation based on the average energy, and determines the division coefficient according to the division coefficient. Since the quantization of 2 is performed, the quality of the reconstructed signal is kept good,
It has the advantage that bit allocation can be done at low cost.

[Brief description of drawings]

FIG. 1 is a block diagram for realizing a threshold value control quantization determination method according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a processing flow of the first embodiment.

FIG. 3 is a flowchart showing a processing flow of a threshold value control quantization determination method according to a second embodiment of the present invention.

FIG. 4 is a flowchart showing a processing flow of a threshold value control quantization determination method according to a third embodiment of the present invention.

FIG. 5 is a diagram showing a comparison result of bit allocation.

FIG. 6 is a block diagram for realizing a conventional quantization decision method.

FIG. 7 is a flowchart showing a processing flow of the conventional quantization determination method of FIG.

FIG. 8 is a block diagram for implementing another conventional quantization decision method.

[Explanation of symbols]

 101 Filter Bank 102 Maximum Value Determining Means 103 First Quantizer 108 Second Quantizer 110 Power Estimating Means 111 Bit Allocation Means

Claims (6)

[Claims] 1. A subbank sample is generated from an input audio signal by a filter bank, a maximum value is obtained from the subband sample, the maximum value is quantized by a first quantizing means, and the quantized maximum value is obtained. Normalize,
In the process of quantizing it by the second quantizing means,
An average energy of the sub-band sample is calculated, a division coefficient is obtained by a predetermined calculation based on the average energy, and the second quantization is performed according to the division coefficient. Threshold control quantization decision method. 2. A component of said frame within a defined time interval in determining the quantization of a digital audio signal with spectral and temporal structure, each represented by a plurality of frames of spectral quantized components. The peak value of the frame, removing the component in the frame by comparing the peak value with the hearing threshold in the human auditory system, and calculating the average energy of the frame within the time interval. A threshold control quantization determination method for an audio signal, comprising: changing a quantization step of the frame according to a predetermined relationship between the step and the average energy. 3. The method of claim 2, wherein the component is set to 0 when the peak value is less than the hearing threshold. 4. The step of quantizing a component is determined by iteratively selecting the frame having the highest average energy and increasing the quantization step while decreasing the average energy non-linearly. A threshold control quantization determination method for an audio signal according to claim 2. 5. A threshold control quantization decision method for an audio signal as claimed in claim 4, characterized in that the quantization is improved when groups of frames are matched by different non-linear reductions of the mean energy. . 6. Before iterating, if each frame first matches, then it becomes more efficient to determine the quantization step, and the step is derived in proportion to the mean energy of the frame. Claim 4 characterized by
A threshold control quantization decision method for an audio signal as described.