KR100827458B1 - Method for audio signal coding - Google Patents

Abstract

The present invention relates to an audio encoding method. An audio encoding method of the present invention is a method of performing quantization and compression encoding based on time / frequency transform by applying a psychoacoustic model during compression encoding of an audio signal, the method comprising: obtaining a masking threshold for each band of an input audio signal; Performing pre-echo control; Obtaining a PE (Perceptual Entropy) value for block switching or window switching; Obtaining a PE (Perceptual Entropy) value for bit allocation per section; Characterized in that it comprises a.

Audio, Coding and Psychoacoustic Models

Description

Audio coding method {METHOD FOR AUDIO SIGNAL CODING}

1 is a diagram showing the structure of an audio encoder to which the present invention is applied.

2 is a flowchart illustrating a preeco adjustment and a masking threshold calculation process in an audio encoding method according to an embodiment of the present invention.

<Explanation of symbols for the main parts of the drawings>

210: MDCT unit 220: FFT unit

230: psychoacoustic model unit 240: window conversion unit

250: quantizer 260: encoder

270: bit string component

The present invention relates to an audio encoding method.

The MPEG audio coding algorithm aims to compress an audio signal without losing subjective sound quality in order to reduce the huge channel capacity required for storing and transmitting the audio signal. For this, we use Perceptual Coding based on human sensory characteristics. Perceptual coding is a method that uses a minimum audible limit, which is the minimum level that can be sensed by hearing, and a masking phenomenon in which other sounds are hard to be heard by a specific sound. The minimum audible limit depends on the frequency of the note (high and low), and the masking phenomenon depends on the masking note (Masker) and the frequency of the masked note (Maskee). In particular, the frequency width at which the masking effect occurs is called a critical band, and the perceptible signal-to-noise ratio (S / N ratio) within the critical band is very low. Accordingly, in MPEG audio coding, compression coding based on the above perceptual coding is performed to mix digital audio signal quantization noise within a critical band so that the quantization noise is not represented.

As described above, MPEG audio uses a lossless compression method together with a statistical lossless compression method for compressing an audio signal, so that the masked phenomenon of psychoacoustic theory is not perceived by the human ear. Therefore, when performing the encoding process, the maximum allowable amount of noise for each frequency is obtained through a complicated process called a psychoacoustic model. This needs to be taken into account, so the role of the psychoacoustic model is very important for obtaining high quality audio output signals.

An object of the present invention is to provide an audio encoding method which increases the efficiency of an encoding process in an audio encoder and enables effective bit allocation for encoding.

An audio encoding method according to the present invention for achieving the above object is a method for performing quantization and compression encoding based on time / frequency conversion by applying a psychoacoustic model during compression encoding of an audio signal, and includes block switching or obtaining a Perceptual Entropy (PE) value for window switching; Determining whether to switch sections according to the PE (Perceptual Entropy) value; Obtaining a PE (Perceptual Entropy) value for bit allocation in each section according to the section switching decision result; Characterized in that it comprises a.

In addition, an audio encoding method according to the present invention for achieving the above object is a method of performing quantization and compression encoding based on time / frequency transform by applying a psychoacoustic model during compression encoding of an audio signal, the band of the input audio signal Obtaining a star masking threshold; Performing pre-echo control; Obtaining a PE (Perceptual Entropy) value for block switching or window switching; Obtaining a PE (Perceptual Entropy) value for bit allocation per section; Characterized in that it comprises a.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention.

The present invention is to improve the efficiency of the bit allocation process of the audio coder using MPEG psychoacoustic model-II, the pre-echo control process in the last step of obtaining the masking threshold in the psychoacoustic model Is divided into long block and short block.

1 shows an embodiment of an audio encoder structure to which the audio encoding method of the present invention is applied. The audio encoder shown in FIG. 1 is an MPEG-based audio encoder. Looking at the configuration, the MDCT unit (Modified Discrete Cosine Transform) 110, the Fast Fourier Transform (FFT) unit 120 of the input audio signal, psychoacoustic model unit 130, window transform unit 140, quantization unit ( 150, an encoder 160, and a bit string generator 170. The quantization unit 150 includes a quantization and bit allocation unit 151 and a Huffman coding unit 152, and the encoding unit 160 includes a temporal noise shaping (TNS) unit and an intensity / coupling unit (Insensity / Coupling). 162, a prediction unit 163, and an M / S unit 164.

As shown in FIG. 1, the input audio signal is converted into a frequency axis signal through the MDCT analysis filter 110 for encoding, and then encoded through various methods. At the same time, the psychoacoustic model unit 130 analyzes the perceptual characteristics of the input signal to determine the maximum allowable quantization noise for each frequency required for the bit allocation process. The bit allocation process optimizes the quantization noise generated in the quantization process at a given bit rate to be as small as possible than the maximum allowable noise obtained from the psychoacoustic model.

The psychoacoustic model requires a frequency conversion process of the input signal because it analyzes the perceptual characteristics of the input signal on the frequency axis. As shown in FIG. 1, the frequency conversion is already performed in the encoding process through the MDCT analysis filter 110, but since the experimental results of the psychoacoustic theory are mostly made on the Discrete Fourier Transform (DFT) axis, the MPEG standard proposes a psychoacoustic model. It is recommended that a separate Fast Fourier Transform (FFT) transform is required.

MPEG psychoacoustic model II has a method that encodes using short-term windows to prevent the occurrence of pre-echo, and the band-specific masking threshold of the current frame is smaller than that of the previous two frames. We recommend a pre-echo control method that applies to the masking threshold of the current frame.

The MPEG standard proposes to determine the block switching decision through the PE (Perceptual Entropy) value, which is one of the psychoacoustic model results. For this purpose, the FFT operation uses two types of windows: long window and short window. In contrast, the MDCT analysis filter 110 further uses a long start window and a long stop window in addition to the long window and the short window.

The psychoacoustic model determines that the next section is a short block, and if block switching occurs from a long section to a short section, the MDCT analysis filter uses the long section start window for the current section. If the short section is to be switched to long section, the long section finishing window is used.

Referring to FIG. 1, an operation of an audio encoder will be described. The MDCT unit 110, as an MDCT analysis filter, converts the input audio signal into a discrete cosine transform (DCT) to convert the input audio signal into a frequency axis. As described above, the FFT unit 120 converts an input audio signal into a frequency axis for psychoacoustic modeling. Here, the input of the MDCT analysis filter and the window form of the FFT operation have the same window form converted by using the time axis information in the window converter 140 of the front end. That is, the long-term start window and the long-end window that are used in the MDCT analysis filter are used for the FFT calculation of the psychoacoustic model.

As described above, the FFT and MDCT are time / frequency conversions. In order to use a characteristic that is easier to encode a signal in the frequency domain than a signal in the time domain, the FFT and MDCT convert the audio signal in the time domain into an audio signal in the frequency domain. In this case, since the length of the conversion window is closely related to the frequency resolution, it is appropriately selected, which is provided from the window conversion unit 140 using time axis information.

The psychoacoustic model unit 130 models human auditory characteristics for perceptual encoding of multi-channel audio. The psychoacoustic model unit 130 extracts the characteristics of the input audio and calculates a degree of quantization noise that is not detected by the human auditory for each band. The allocation of bits reflects this to achieve optimal coding. The techniques and implementation of psychoacoustic modeling are the same as those used in the existing psychoacoustic modeling based audio coding algorithm.

The quantization unit 150 performs audio signal quantization based on a method of allocating an optimal quantization level for a given bit rate by using the psychoacoustic model unit 130 on the frequency spectrum compressed by the encoder 160. This is performed by the quantization and bit allocation unit 151, and the quantized frequency spectrums are composed of values represented by allocated bits, which can restore the original values at the decoder to represent them with fewer bits. As a method of encoding in a state, for example, the Huffman coding unit 152 uses a technique of encoding a reduced number of bits using Huffman coding.

The encoder 160 performs compression encoding on the audio signal by using methods that can reduce or predict the amplitude of the frequency spectrum provided by the time / frequency converter-MDCT unit 110 for compression encoding the audio signal. . For this purpose, the TNS unit 161, the strength / combination unit 162, the prediction unit 163, and the M / S unit 164 are used.

The TNS unit 161 minimizes quantization noise by predictively coding noise generated in the quantization process in a frequency domain. Intensity / combination unit for a technique to reduce the amount of data actually transmitted by transmitting only the level difference of another channel for one channel for each pair of channels divided as left and right channels by the relationship between channels. / Coupling) 162 to perform encoding. In addition, as a data compression method in the time domain, a prediction unit 163 is used for inter-frame prediction for predicting a spectrum of a current frame from a spectrum of a previous audio frame, which is transmitted by transmitting only a prediction parameter and a prediction error. It provides a basis to reduce the amount of data. The left and right channel signals are converted into M (Middle) / S (Side) channels to encode the data using the M / S unit 164 for reducing data. The TNS, Intensity / Coupling, Prediction, and M / S processes are selectively used encoding processes used to increase coding efficiency, and Huffman coding is a lossless coding process used to encode quantized spectral information.

The bit stream constructing unit 170 generates a bit stream of the compressed and encoded audio data. That is, the header information and the spectral data of the bit string are configured as bit strings. In this case, an audio element stream (ES) is converted into a packetized element stream (PES), which is a packetized bit string, according to external control or user control. It may include.

As mentioned above, the MPEG standard allows the block switching or window switching decision to be determined through the PE (Perceptual Entropy) value, which is one of the results of the psychoacoustic model. The PE value may be expressed as a sum of signal-to-mask ratio (signal-to-mask ratio) for each band. When a sudden increase in signal occurs on the time axis, energy increases across the entire band on the frequency axis. This also increases the masking threshold across the band. This increased masking threshold is compared with the masking threshold of the previous frame during the pre-eco adjustment process, and if the smaller value is used as the masking threshold of the current frame, the SMR in each band of the current frame is increased so that the PE value is increased. do. If the PE value exceeds the predefined threshold, the coding is performed by switching from the long section to the short section.

However, the PE value for block switching is calculated using the energy and masking threshold in the long term, while the pre-eco phenomenon occurs in the short term. Therefore, even when the current frame is determined to be long, the masking threshold is lower than the actual predicted value by using the pre-eco adjustment method, so that more bits are allocated to these bands than are actually needed in the bit allocation. This reduces the bits to use in relatively other bands. Therefore, quantization noise in these bands can be increased at low bit rates.

In other words, in order to prevent the pre-echo, when the long section is encoded by dividing the long section, although the pre-eco is not generated in the long section, the masking threshold in the long section is compared with the values of the previous frame. The long section is encoded using a masking threshold smaller than the calculated value. This approach lowers the masking threshold of bands in the long term, allowing more bits to be used than are actually needed. This reduces the bits used in other bands at lower bit rates. Therefore, quantization noise is increased.

The present invention provides a method for effective bit allocation in the long term in MPEG psychoacoustic model II. For this purpose, the PE value for interval switching and the PE value for bit allocation are calculated separately in psychoacoustic model II. Equation 1 is an equation for obtaining a PE value.

Here, b is a band unit, and (w_high [b] -w_low [b]) is the number of fast Fourier transform coefficients in one band. thr [b] and e [b] represent the masking threshold and energy of each band. According to Equation 1, the PE value is the number of FFT coefficients (w_high [b] -w_low [b]) obtained from the result of performing the FFT on the input audio signal, the masking threshold (thr [b]) for each band, and It can be seen that it can be obtained from the sum of the energy of each band (e [b]).

Thus, the masking threshold thr [b] for each band for obtaining the PE value is obtained for the long and short sections. The band-specific masking threshold thr [b] is obtained by using an absolute audible threshold value, a masking threshold value of a previous frame and a previous frame, and a predefined weight. Equation 2 below is a formula for calculating a masking threshold to be used to calculate the PE value for the interval switching.

Here, absthr [b], nb_l [b], and nb_ll [b] denote absolute threshold values and masking thresholds of the previous frame and the previous frame, respectively, and rpelev and rpelev2 are predefined weights. nb [b] is a masking threshold calculated in the current frame until the pre-eco adjustment process, and the PE value for the interval switching is calculated using thr [b].

As described above, in the audio encoding method of the present invention, the interval switching PE value and the bit allocation PE value are separately calculated. The threshold to be used to calculate the PE value for bit allocation in the long term is calculated as in Equation 3 below.

Here, the masking threshold thr [b] for each band can be obtained by using the absolute audible threshold absthr [b] and the masking threshold nb [b] calculated in the current frame until the pre-echo adjustment process, and this threshold thr [b] We can calculate the PE value for bit allocation in the long term by using. That is, the masking threshold in the long term is obtained by comparing only the absolute audible threshold result value without comparing the masking threshold with the previous frame.

On the other hand, the masking threshold in the short section uses the equation (2) to prevent the pre-eco generation. Therefore, both the masking threshold for calculating the PE value for the interval switching and the masking threshold in the short section are calculated using Equation 2.

In the audio encoding process according to the embodiment of the present invention, as shown in FIG. 1, audio data is input, MDCT analysis is performed by applying a window, FFT transform for psychoacoustic modeling, Modeling of the audio signal based on the determination and control of the quantization bit number allocation, and performing compression coding of the audio signal based on the output of the MDCT unit, and based on the psychoacoustic modeling results of allocation of the quantization bit number and Huffman coding The quantization of the audio signal is performed, and a bit string of compressed and encoded audio data is output.

FIG. 2 is a flowchart illustrating a process of calculating pre-eco adjustment and masking threshold of MPEG psychoacoustic model II in the audio encoding process.

The first step S10 is a step of calculating a masking threshold thr [b] for each long and short periods. The masking threshold thr [b] obtained in this step is a masking threshold value which has not undergone the pre-eco adjustment process and is used in the PE value calculation step S51 for bit allocation in the long term.

The second step S20 is a step of performing pre-eco adjustment on the long and short sections based on the masking threshold comparison with the previous frames. By performing this step, the masking threshold thr [b] subjected to the pre-eco adjustment process is used in the PE value calculation step S61 for bit allocation in the short term.

The third step (S30) is a step of calculating the PE value for section switching (long section => short section) in the long section. The calculation for obtaining the PE value is performed according to Equation 1, and the masking threshold to be used to calculate the PE value for the interval switching is calculated according to Equation 2.

A fourth step S40 is a step of comparing the calculated PE value with a predefined value SWITCH_PE for the interval switching. If the calculated PE value does not exceed the predefined SWITCH_PE value, the process proceeds to the fifth step S51 to perform encoding in the long term. If the calculated PE value exceeds the predefined SWITCH_PE value, the eighth step S61 is performed. Transitioning from the long section to the short section, the encoding is performed.

The fifth step (S51) is a step of calculating the PE value for bit allocation in the long section using the masking threshold that has not undergone the pre-eco adjustment process. The calculation for obtaining the PE value is according to Equation 1, and the masking threshold to be used to calculate the PE value for bit allocation in the long term is calculated according to Equation 3.

The sixth step S52 is a step of obtaining a masking threshold for each scale factor band in the long term, and the next seventh step S53 is an SMR (ratio of signal to masking thresholds) for each scale factor band in the long term. Step to calculate.

An eighth step S61 is a step of calculating a PE value for bit allocation in a short section by using a masking threshold that has undergone pre-eco adjustment according to the section switching. Here, the calculation for obtaining the PE value is performed according to Equation 1, and the masking threshold to be used to calculate the PE value for bit allocation in the short term is calculated according to Equation 2.

A ninth step S62 is a step of obtaining a masking threshold for each scale factor band in a short section, and a next tenth step S63 is a scale factor band SMR (ratio of a signal to masking threshold) in a short section. Step to calculate.

As described above, in order to obtain PE values for bit allocation and interval switching in psychoacoustic model II, masking threshold values in the long section are used before and after the pre-eco adjustment, respectively. For this, the long section can be used separately before and after the pre-eco adjustment process. In the case where a method other than the PE value is used for the interval switching, the interval switching part may be replaced by these other methods.

According to the audio encoding method of the present invention, the pre-eco adjustment process is used only in a short section to reduce the occurrence of pre-eco in the short section, as well as to use the masking threshold predicted in the current frame in the long section. Bit allocation can be effectively made.

Claims (9)

A method of performing quantization and compression encoding based on time / frequency transform by applying a psychoacoustic model during compression encoding of an audio signal, Obtaining a PE (Perceptual Entropy) value for block switching or window switching; Determining whether to switch sections according to the PE (Perceptual Entropy) value; Obtaining a PE (Perceptual Entropy) value for bit allocation in each section according to the section switching decision result; Audio encoding method comprising a. The audio encoding method of claim 1, wherein the PE (Perceptual Entropy) value is obtained by using an FFT coefficient, a masking threshold for each band, and an energy value in one band. A method of performing quantization and compression encoding based on time / frequency transform by applying a psychoacoustic model during compression encoding of an audio signal, Obtaining a band-specific masking threshold of the input audio signal; Performing pre-echo control; Obtaining a PE (Perceptual Entropy) value for block switching or window switching; Obtaining a PE (Perceptual Entropy) value for bit allocation per section; Audio encoding method comprising a. The audio encoding method of claim 3, wherein the PE (Perceptual Entropy) value is obtained by using the number of FFT coefficients in one band, a masking threshold for each band, and energy. 4. The audio encoding method of claim 3, wherein the PE (Perceptual Entropy) value for the interval switching and the PE (Perceptual Entropy) value for the bit allocation are obtained based on different masking thresholds. The audio encoding method of claim 3, wherein the PE (Perceptual Entropy) value for bit allocation is obtained for each section. 4. The audio encoding method of claim 3, wherein a masking threshold for obtaining a PE (Perceptual Entropy) value for bit allocation uses a value before and after the pre-eco adjustment according to a section. 8. The audio encoding method of claim 7, wherein a masking threshold for obtaining a PE (Perceptual Entropy) value for bit allocation uses a value before pre-eco adjustment in a long period. The audio encoding method of claim 7, wherein a masking threshold for obtaining a PE (Perceptual Entropy) value for bit allocation uses a value after pre-eco adjustment in a short period.

Images