KR100640833B1 - Method for encording digital audio - Google Patents

Abstract

A method for encoding digital audio is provided to reduce the number of repetition of quantization loop in which non-linear quantization and Huffman encoding are performed. A quantization loop performing process in a method for encoding digital audio includes the steps of: receiving the number of used bits and the size of early quant(300); non-linear quantizing an MDCT coefficient for the size of input quant, and encoding the quantized coefficients and calculating the number of used bits(310); and judging whether the number of bits calculated at the previous step is smaller than the number of usable bits at this step and larger than the number of bits subtracting an offset value calculated in a predetermined linear function from the number of usable bits(320).

Description

Method for encoding digital audio {Method for encording digital audio}

1 is a flowchart of an audio encoding process in MP3 format

2 is a flowchart of an audio encoding process in AAC format

3 is a flowchart illustrating a dual structure for bit allocation and quantization in an audio encoding process

4 is a flowchart illustrating a method of calculating a conventional quantization loop.

5 is an experimental result graph showing the relationship between the quantization step size and the number of bits used in the quantization loop.

6 is a flowchart illustrating a method of calculating a quantization loop according to the present invention.

The present invention relates to a method of encoding digital audio, and more particularly, a computational amount and execution time of an iterative loop process for bit allocation and quantization, which require a large amount of computation during the encoding process of digital audio, thereby making real time processing and low power design difficult. The present invention relates to a digital audio encoding method which can greatly reduce the number of times.

Digital audio, which has much cleaner sound quality and wider bandwidth than conventional analog audio, has been applied to digital TV, digital multimedia broadcasting (DMB) and digital audio broadcasting (DAB) as well as CD and DVD. There is a growing interest. Digital audio can also be output in a variety of media, such as movies and computers, and is expected to continue to expand its application. However, when the analog audio signal is sampled to obtain a high quality digital audio signal, a large amount of data increases during processing, which takes up a lot of memory and a transmission band. In particular, when using transmission channels such as digital TV, DMB, DAB, and network communication network, real-time transmission is impossible without effectively compressing a large amount of data.

Audio coding, which effectively compresses digital audio signals, has been developed by AT & T, Dolby Labs in the US and IRT, Philips, CCETT in Europe, and Sony in Japan since the 1980s. Among the audio compression technologies currently used worldwide, MPEG based on MUSICAM (Masking pattern adapted Universal Subband Integrated Coding And Multiplexing) and ASPEC method, and AC-3 method proposed by Dolby Lab of USA. High quality audio coding algorithm should provide CD quality high quality sound after encoding and decoding audio signal and should not have perceptual loss compared with original sound. These coding algorithms commonly combine a compression method using the auditory characteristics of the human ear to remove perceptual redundancy of a signal and a conventional data compression method to remove statistical redundancy of a signal.

The third layer audio encoding algorithm (MP3) of MPEG-1,2 is a high quality audio encoding technology that can reproduce the original sound and the acoustically insensitive reconstructed sound even when the CD level audio signal is compressed by 10: 1 or more. In year (MPEG-1) and 1994 (MPEG-2), international standards were completed respectively. In particular, as MP3 is widely used for audio playback using the Internet, the performance is recognized and the market demand is rapidly increasing. As a portable MP3 player is developed, it is widely used as a portable digital audio device replacing a CD player and a cassette player.

In the case of the MP3 decoding algorithm, the operation process is relatively simpler than that of the encoding algorithm, and many low-power processors and ASIC chips capable of real-time processing have been developed and used. However, in the case of an encoder requiring a relatively large amount of computation, It is not. However, as the portable MP3 market expands and diversifies, there is an increasing demand for high-quality recording, that is, high-quality audio encoding in a portable audio device. In particular, the most widely used MP3 encoding can be said to be very useful.

On the other hand, MPEG-2,4 AAC (Advanced Audio Coding) is MPEG's high-quality audio coding method developed in 1997 by overcoming the limitations of the existing MP3 coding method and incorporating newly developed audio compression signal processing technology. The standardization was completed under the name of MPEG-2 AAC, and it is used as it is in MPEG-4. MPEG-2,4 AAC provides high compression efficiency and high sound quality compared to MP3 and is used as audio coding technology for broadcasting and communication. Since the AAC fully accommodates the advantages of the MP3 encoding method, the AAC includes many similar processes to the MP3 in the encoding and decoding process. In addition, since the encoding process requires much more computation than the decoding process, there are many difficulties in the low power design of the encoder.

The encoding process of the MP3 and AAC is more complicated than the decoding process because active bit allocation by the psychoacoustic model occupies a large amount of computation. The psychoacoustic model includes many exponential and logarithmic operations, which are inefficient to implement with a digital signal processor (DSP), and the bit allocation process performed on the basis of this process is performed until the optimal noise shaping is achieved. In turn, nonlinear quantization and Huffman coding have to be performed repeatedly, resulting in a large execution time. In MP3 and AAC, the operation amount of the psychoacoustic model calculation and bit allocation portion is usually about 80% of the total calculation amount. Especially, the bit allocation process varies greatly depending on the characteristics of the input data. Comes with many difficulties.

Referring to Figure 1, the basic structure of a conventional MP3 audio encoding algorithm will be described. In MP3, first, the input signal is passed through 32 weighted superposition-added equally spaced filter banks to convert to subband samples to remove statistical redundancy of the audio signal. By performing a transform, a hybrid transform encoding method for increasing frequency resolution is used. At the same time, an FFT (Fast Fourier Transform) is performed on the audio signal, and a masking threshold, which is an inaudible noise level, is obtained to remove perceptual redundancy due to human auditory characteristics in the psychoacoustic model. Based on the masking threshold in each frame, the quantization bits of the MDCT spectrum are allocated so that the quantization noise can be masked by the signal.The bit allocation and quantization (step 121) and the Huffman coding process (step 125) are repeated loops. Is processed through (100). When the noise cannot be completely masked, bits are allocated per scale factor band to minimize subjective noise, and the quantized MDCT spectrum is Huffman coded and encoded together with additional information (step 190) to form a bit string.

Referring to Figure 2 describes the basic structure of the conventional AAC coding algorithm as follows. The masking threshold is obtained in the frequency domain through the psychoacoustic model, and the structure of performing bit allocation to shape the noise so that the quantization noise is inaudible is the same as that of the MP3. However, AAC encoding performs MDCT filter bank on 2048 points with higher frequency resolution compared to MP3, and uses temporal noise shaping (TNS), a technique for shaping noise in the time domain. The use of prediction to remove statistical redundancy between them is very different from MP3. The AAC encoder also includes a preprocessor that can vary the sampling frequency according to the bit rate to be transmitted.

In MP3, a series of processing processes as shown in FIG. 1 are performed in units corresponding to 576 samples in the time domain, which is called granules. The two granules form one frame. The frame becomes a minimum unit capable of independently decoding. Meanwhile, in the AAC, the process of FIG. 2 is performed in units corresponding to 1024 samples in the time domain, which is called a frame. Therefore, the frame in MP3 and the frame in AAC have different lengths and meanings. In the present specification, for convenience of description, the term 'frame' will be referred to collectively. Therefore, the frame unit referred to in the present specification means a granule unit for MP3.

Referring to FIG. 3, a process of performing quantization and Huffman coding through bit allocation in the iterative loop included in FIGS. 1 and 2 will now be described. The iterative loop compares the quantization noise and the masking threshold to determine the distortion allocation loop (outer loop) (step 100) and bit allocation results and number of bits available for quantization (step 121) and encoding (step 125). It comprises a dual structure of a quantization loop (inner loop) (step 120), which proceeds in comparison (called a nested loop). Inputs of the iterative loop include MDCT coefficients as quantization targets, masking thresholds as outputs of the psychoacoustic model, and the number of available bits b _a determined by the compression bit rate. First of all, the inner quantization loop performs nonlinear quantization of MDCT coefficients. The equation for nonlinear quantization is as follows.

In Equation 1, n is an MDCT coefficient index of 0, 1, 2, ..., N-1, where N is 576 for MP3 and 1024 for AAC. xr (n) is an MDCT coefficient having a real value, and is (n) is a quantized MDCT coefficient. In Equation 1, quant is actually called a quantization step size or global gain. The actual quant value determines the number of bits used for encoding the frame and the quantization noise, so it is the value to be controlled in the quantization loop.

Huffman coding is performed on the quantized MDCT coefficients (step 125) through a nonlinear quantization process (step 121), through which the corresponding frame is encoded and the number of bits b _u used is calculated. If the calculated value b _u is greater than the number of available bits b _a , the number of bits is exceeded. Therefore, the quantization step size should be adjusted to perform quantization and Huffman coding again. 120 may escape (step 127). When quantization and Huffman coding are performed through an internal quantization loop, the external distortion control loop performs inverse quantization from quantized MDCT coefficients, and distortion, that is, quantization, from the difference between the dequantized MDCT coefficients and the original real MDCT coefficients. Calculate the noise value. The calculated distortion is compared with the masking threshold, which is the output of the psychoacoustic model, in units of a frequency domain called a scale factor band (step 155), and the comparison results in the case where the distortion value is larger than the masking threshold (this distortion is It is perceived as annoying sound to the human ear) and the scale factor value for the band is updated (step 157) and returned to the internal quantization loop. In the inner loop, the iterative loop performs nonlinear quantization and Huffman coding on the MDCT coefficients xr (n) values newly determined by the scale factor. Through this, if the conditions for distortion and masking threshold for all bands are satisfied, the entire iteration loop is completed. As described above, since many iterations are performed until the exit condition of the outer loop and the exit condition of the inner loop are satisfied, the iteration loop for bit allocation and quantization has a large amount of computation.

The reason that the quantization and Huffman coding must be repeated for each loop in the internal quantization loop is because of Huffman coding. Huffman coding means that the length (the number of bits) varies depending on the probability of generating the data (symbol) to be encoded. Encoding using a codeword is a method. Therefore, the actual number of bits used is not determined by factors such as the quantization step, but can be known only after performing quantization and Huffman coding. Therefore, the quantization and Huffman coding must be repeated several times until an optimal value is obtained. Therefore, there is a problem that requires a large amount of computation to process the process.

Referring to FIG. 4, the quantization loop algorithm is described in terms of the quantization step size s (i) to be controlled and the number of bits b _a used as follows. In the quantization loop, the number of bits b _a available in the current frame and the initial value s (0) for the quantization step quant are input (step 200). In each loop, the number of bits b _u required to quantize MDCT coefficients is calculated through a process represented by a function B (*) (210), and b _u has the following relationship with the function B (*).

Where i is the index of the loop. Function B (*) is actually a series of processes that nonlinear quantize MDCT coefficients using the current quantization step value s (i), perform Huffman coding on the quantized coefficients, and then calculate the number of bits used. . The function value B (*) is input to the used number of bits b _u (step 210). In FIG. 4, the quantization loop process is repeated until condition Cnd-I (step 220) and condition Cnd-II (step 240) are satisfied. The conditions Cnd-I and Cnd-II are as follows.

That is, the condition Cnd-I of Equation 3 checks whether the value b _u (i) calculated in the current loop is less than or equal to the available number of bits b _a (step 220). If this condition is not satisfied, the number of bits available in the current frame is exceeded, and the quantization step size is adjusted by the function UPD _I (*) (step 230), and the loop is performed again. When the condition Cnd-I is satisfied, the condition Cnd-II performs a B (*) function on the quantization step s (i) -1 which is one smaller than the current quantization step size that satisfies the condition Cnd-I. The output b _u value is compared with ba to determine whether it is large (step 240). Condition Cnd-II is performed to determine whether the quantization step satisfying the condition Cnd-I is an optimal value. If the condition Cnd-II is not satisfied, the quantization step size is adjusted by the function UPD _II (*) (step 250), and the loop is performed again. By looping until both conditions Cnd-I and Cnd-II are satisfied, we finally get outputs such as the optimal quantization step size value, the number of bits used, the actual quantized MDCT coefficients, and the Huffman codebook index used. Can be.

Observing the above-described process of the quantization loop, it can be seen that the number of iterations of the loop depends on the performance of the functions UPD _I and UPD _II updating the initial value s (0) for the quantization step and the quantization step size.

The most basic implementation method according to the ISO / IEC 11172-3 MPEG standard is to set the initial value s (0) of the quantization step size to a minimum value and the update function UPD _I as follows.

That is, it increments by one from the smallest quantization step size and loops until the condition Cnd-I is satisfied. Since the quantization step is increased by one step from the smallest value, when the condition Cnd-I is satisfied, the condition Cnd-II is automatically satisfied. While this is the surest way to find the optimal value most easily, it is very inefficient in terms of real-time processing. If the final quantization step value to be obtained is a very large value, the number of loops increases proportionally, which results in an increase in the amount of computation. For example, in the case of MP3, since the quantization step can have a value from 0 to 255, in the worst case, 256 loops should be performed, which is an amount of computation that is impossible in real time.

One possible efficient method is to use the quantization step size for the previous frame as the initial value s (0). This uses the property that the quantization step size between neighboring frames does not change significantly, and can have a great effect on a section in which such a property is maintained. However, if the magnitude of the audio signal suddenly changes greatly, the quantization step size also changes abruptly. That is, the number of repetitions of the quantization loop also increases in proportion, thus increasing the amount of computation. In the real-time audio coding system, since the computational capability of the system is determined based on the maximum computational demand, not the average computational amount, this method has been improved in the average sense, but there is a problem in that the worst case has no improvement. .

In addition, when the amount of computation increases due to the increased number of iterations of the quantization loop, a problem arises in that a lot of power is consumed in the encoding process of the digital audio.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to encode digital audio, which can greatly reduce the amount of computation for a quantization loop in a repeating loop for bit allocation, which requires a large amount of computation during the encoding process of digital audio. To provide a way.

Another object of the present invention is to provide a new digital audio encoding method which reduces the computational amount of the audio encoding and reduces the power consumption of the audio encoding apparatus capable of performing the above process but does not affect the sound quality.

In order to achieve the above object, the present invention provides a coding method of digital audio including a step of performing a quantization loop for bit allocation and quantization of an audio signal, wherein the step of performing the quantization loop is (a) the number of bits used and the initial stage. Receiving a quantization step size; (b) nonlinear quantizing the MDCT coefficients with respect to the input quantization step size, and performing encoding on the quantized coefficients to calculate the number of bits used; (c) It is determined whether the number of bits calculated in the step (b) is less than or equal to the number of bits available in the loop step, and the number of bits obtained by subtracting the offset value obtained in a predetermined first function relation from the available number of bits is exceeded. Doing; And (d) exiting the quantization loop if the number of bits calculated in step (b) satisfies the above condition, otherwise updating the quantization step size of the next loop. After performing the step of returning to step (b) it provides a digital audio encoding method comprising the step of.

The linear function of step (c) is a function approximating the relationship of the number of bits used to the size of the quantization step. And, the offset value of step (c) may be arbitrarily set, but it is preferable to set the change value of the first function with respect to the change of the arbitrary quantization step size as an offset value.

In the present invention, the updating of the quantization step size of the loop immediately following the loop of step (d) includes: (d1) multiplying the offset value by a factor of 1 or less from the number of bits used in step (a). Subtracting to estimate the number of bits to be used in the next loop; And (d2) the first step of step (c) with respect to the number of bits used in the loop from the result of the inverse of the first function of step (c) with respect to the number of bits to be used in the next loop estimated in step (d1). And adding the value obtained by subtracting the result of the inverse function of the function to the quantization step size of the loop to update the quantization step size of the next loop.

In addition, it is preferable that the size of the initial quantization step received in the step (a) is the final quantization step size value of the previous frame when the loop is not the maximum loop. When the magnitude of the audio signal between frames does not change rapidly, the number of quantization loops can be further reduced by using the final quantization step size of the previous frame as the size of the initial quantization code step.

 When the quantization step size and the number of bits used are approximated to the first function, the number of bits to be used in the next loop can be estimated from the number of used bits of the loop from the relationship of the first function. Therefore, the present invention can estimate the number of bits to be used by a constant relation without increasing the loop index by one as in the related art, so that the quantized loop escape condition can be satisfied quickly. Thus, the computation amount and time for computing the quantization loop can be significantly reduced.

A method of encoding digital audio according to the present invention will now be described with reference to FIG. 5.

In the method of encoding digital audio according to the present invention, the quantization loop method is based on a statistical relationship between s (i) (quantization step size), which is the input of function B (*), and output b _u (i) (number of bits used). depart. 5 is a graph showing experimental results showing the relationship between the input s (i) and the output b _u (i) of the function B (*). In order to obtain the graph of FIG. 5, the input value s (i) is increased from the minimum value to the maximum value and substituted into the function B (*) to obtain the output b _u (i) for the process of the corresponding inner quantization loop. The points corresponding to the input and output pairs for various input audio signals are displayed on a graph. The points shown in FIG. 5 show the results for some of the results for the various audio signals. As can be seen in FIG. 5, the relationship between s (i) and b _u (i) is substantially linear for each section (linear relationship is indicated by a thick line in FIG. 5). This means that although the actual function B (*) has a series of unpredictable nonlinear characteristics with respect to the number of bits used, the result is generally linear. Based on these linear characteristics, the following relation can be established.

In Equation 6, c _k is an unknown constant, which is an unknown value that can be fixed in the k-th frame. And a is a slope having a positive value of Equation 6, which is linear. With respect to Equation 6, the relation in the i + 1 iteration loop can be expressed as the following equation.

The following Equation 8 can be derived from Equations 6 and 7 below.

Equation 8 uses the linear relationship of FIG. 5 to input s (i + for use in the next loop from s (i), b _u (i) of the current loop and the expected output b _u (i + 1) in the next loop. 1) can be obtained. In Equation 8, if s (i + 1) can be predicted by the value of b _u (i + 1), the quantization step size of the i + 1 loop is equal to the quantization step size of the i loop, and the value of b _u (i) and b. _u (i + 1), because each of the values can also be expressed as the inverse of the sum of the difference between the value of the linear function b (*) 5 (a linear function is shown in b _u values for the s value, b _u value If we know the change of, we can get the change of s value from the inverse relation of the linear function),

That is, by replacing b _u (i + 1) with the expected number of bits b _d in Equation 8, Equation 9 can be obtained.

In Equation 9, the slope variable a is a value that can be found through an experiment, and may have a different value depending on a section in which the current s (i) value exists, and in MP3, a value between 50 and 200 is selected. It is desirable to. The b _d is a value obtained by estimating the number of bits to be obtained when the quantization is performed by using the maximum number of bits b _a available in the current frame in the i + 1 loop, and may be calculated as shown in Equation 10.

Since the B (*) function shown in FIG. 5 is estimated in a linear relationship, b _d in Equation 10 means that it can be linearly predicted from the maximum number of available bits b _a for the quantization step size of the i loop. . That is, the value of b _d of the i + 1 loop may be predicted from the value of the i loop b _a at regular intervals (offset in Equation 10) along the assumed linear function of the function B (*). Equation 0.5 is an example and is preferably estimated by multiplying a factor of 1 or less.

Thus, the update to the quantization step is to be obtained as the ratio of the linear distance between the number of bits b _u (i) used by the current step size s (i) and the number of bits b _{d the} output is expected in the next loop. Therefore, the optimum value of b _d can be found actively. This also means that by adjusting a corresponding to the slope variable of Equation 9, the maximum number of iterations of the quantization loop can be adjusted, and thus it can be flexibly applied according to the allowed computation amount.

In conjunction with the quantization step update method described above, the present invention provides a new escape condition of the angularization loop. The new loop escape condition is to avoid the problem of actually increasing the number of executions of the function B (*) to confirm the condition Cnd-II in FIG. 4, and when a bu (i) value satisfying the following condition is obtained, , Exit the quantization loop.

That is, Equation 11 sets the lower limit condition of the value b _u by a constant called offset. If Equation 11 is satisfied, the conventional loop escape condition Cnd-II is satisfied, and an offset value can be set in advance. If the offset value of Equation 11 is determined according to the inverse function of the first order function of FIG. 5, the magnitude of the offset value will vary according to the inclination variable a of Equation 9. In this case, the maximum number of repetitions of the quantization loop may be controlled by adjusting the slope variable a and the offset value of Equation 11. In addition, instead of estimating the offset value as the inverse function of the primary function of FIG. 5, other values may be used depending on b _a and bit rate.

Referring to FIG. 6, an algorithm of an improved quantization loop of a coding method of digital audio according to the present invention will be described.

The number of bits ba available in the current frame and the initial value s (0) for the quantization step quant are input to the quantization loop (step 300). In each loop, the number of bits bu needed to quantize MDCT coefficients is calculated through a process represented by function B (*) (step 310). Then, it is determined whether the condition Cnd-prop expressed in Equation 11 is satisfied (step 320). If the determination condition is not satisfied, the magnitude of the quantization step of the i + 1 loop is calculated by Equation 9 (step 330). Then, the steps 310, 320, and 330 are repeated until the condition 320 is satisfied again. In order to further reduce the number of repetitions, it is preferable that the initial value s (0) of the quantization step size uses the final quantization step size value of the previous frame.

When the digital audio coding method according to the present invention is used, the quantization loop can be performed with an average of 1.8 times and a maximum of 4 times of repetition, so that a maximum of 10 minutes is required compared to a case where a conventional average of 3.9 times and a maximum of 40 times or more are required. An audio encoding process can be implemented with a small amount of computation of 1. In the digital audio encoding method according to the present invention, the quantization step update method in the iterative loop and the escape condition of the quantization loop can be applied to any audio encoding apparatus in which this technique is used.

The effects of the encoding method of digital audio according to the present invention described above are as follows.

First, the encoding method of digital audio according to the present invention can significantly reduce the number of iterations of a quantization loop in which nonlinear quantization and Huffman encoding are performed during the encoding process. Therefore, it is possible to greatly improve the conventional problem that the number of iterations of the quantization loop is increased and the amount of computation is greatly increased.

Second, the digital audio encoding method according to the present invention can expect a more than half reduction effect on the total amount of computation compared to the conventional method, and thus it is possible to design a low power of the real-time digital audio encoding apparatus.

Third, the quantization loop of the digital audio encoding method according to the present invention has little effect on the sound quality performance of the digital audio encoder, and the maximum number of repetitions and the sound quality can be adjusted by adjusting the related factors.

Claims (5)

In the encoding method of digital audio, the process of performing quantization loop for bit allocation and quantization of an audio signal, (a) receiving a number of bits used and an initial quantization step size; (b) nonlinear quantizing the MDCT coefficients with respect to the input quantization step size, and performing encoding on the quantized coefficients to calculate the number of bits used; (c) It is determined whether the number of bits calculated in the step (b) is less than or equal to the number of bits available in the loop step, and the number of bits obtained by subtracting the offset value obtained in a predetermined first function relation from the available number of bits is exceeded. Doing; And (d) exiting the quantization loop if the number of bits calculated in step (b) satisfies the above condition; otherwise, updating the quantization step size of the next loop. And then returning to the step (b). The method of claim 1, And (c) the primary function is a linear function that approximates the relationship between the number of bits used and the size of the quantization step. The method of claim 2, And the offset value of step (c) is a change value of the first function with respect to a change in an arbitrary quantization step size as an offset value. The method of claim 3, wherein Updating the quantization step size of the loop immediately following the loop of step (d), (d1) estimating the number of bits to be used in the next loop by subtracting the offset value multiplied by a factor of 1 or less from the number of bits used in step (a); And (d2) the first function of step (c) with respect to the number of bits used in the loop from the result of the inverse of the first function of step (c) with respect to the number of bits to be used in the next loop estimated in step (d1) And updating the quantization step size of the next loop by adding the value obtained by subtracting the result of the inverse function of the loop to the quantization step size of the loop. The method of claim 1, And the initial quantization step size received in step (a) is input of a final quantization step size value of a previous frame when the loop is not the first loop.

Images