CN101930746B - An Adaptive Noise Reduction Method for MP3 Compressed Domain Audio - Google Patents

Abstract

The invention relates to an MP3 compressed domain audio adaptive noise reduction method. The method is directly based on the MP3 compressed domain for noise reduction. First, MDCT coefficients are extracted from MP3 audio data containing noise, and activity detection is performed on the MP3 audio based on the MDCT spectrum energy characteristics to distinguish active audio segments and silent segments. At the same time, after extracting the MDCT coefficients from the MP3 compressed audio data, according to the sparse characteristics of the MDCT coefficients, a priori statistical modeling is performed on the MDCT coefficients using a normal inverse Gaussian (NIG) distribution function. Then, according to the Bayesian theory, a maximum a posteriori probability estimator based on the NIG prior probability model is designed to obtain an attenuation factor of the corresponding audio segment. In the attenuation noise part, the attenuation factor is used to attenuate the noise of the audio segment, and the number of attenuation iterations is adaptively adjusted according to the attenuation weight of the silent segment audio to achieve noise reduction. Experimental results show that the noise reduction algorithm of the invention can effectively remove the noise in the MP3 audio, improve the signal-to-noise ratio of the compressed audio, and the MP3 audio quality after noise reduction is good.

Description

An Adaptive Noise Reduction Method for MP3 Compressed Domain Audio

technical field

The invention relates to an audio adaptive noise reduction method in an MP3 compressed domain, mainly for realizing the adaptive noise reduction processing method for the MP3 audio containing noise directly in the MP3 compressed domain under different Gaussian white noise conditions.

Background technique

Audio noise reduction technology refers to the use of signal processing and pattern recognition methods to remove noise from audio containing noise, so that the audio after noise removal has a higher signal-to-noise ratio and better quality. Audio noise reduction is one of the key technologies to be solved in the field of audio signal processing.

A large amount of audio data in the Internet and various databases are stored in compressed format. How to process audio data in the compressed domain has become a hot spot in the field of audio research. Scholars at home and abroad have done a lot of research on the segmentation, classification, and retrieval algorithms of compressed audio, and can obtain experimental results similar to those of uncompressed audio processing. However, when the compressed audio is mixed with noise, the accuracy of the audio classification retrieval algorithm is seriously affected. Usually, it takes a lot of time to decompress the compressed domain audio containing noise first, and then perform noise reduction processing, which will inevitably reduce the efficiency of various processing of compressed audio. Therefore, it is particularly important to study how to implement audio noise reduction processing directly based on the compressed domain, and achieve noise reduction with the minimum computational cost to improve the retrieval efficiency of compressed domain audio.

The auditory masking characteristic of the human ear is taken into consideration when the audio is compressed, and the window function of the Modified Discrete Cosine Transform (MDCT) is selected through the second psychoacoustic model. At the same time, the MDCT transform is also a modification of the FFT transform, and the MDCT coefficients are sparse. Therefore, we can extract MDCT coefficients from compressed domain audio, and then try to find a model function that can fit a sparse distribution for prior modeling of MDCT coefficients, and then construct filters to achieve noise reduction for compressed domain audio deal with. The present invention just adopts above-mentioned method, extracts MDCT coefficient from MPEG1 standard audio third-layer compression technology MP3 compression domain audio frequency, adopts normal anti-Gaussian function to carry out prior modeling to the distribution of MDCT coefficient, constructs maximum posteriori probability estimation Function that implements noise reduction for compressed-domain audio.

The noise reduction method proposed by the invention solves the audio noise reduction problem containing noise in the MP3 compression domain, and can be further applied to the speech recognition and classification retrieval system of the MP3 audio.

Contents of the invention

The object of the present invention is to provide a kind of MP3 compressed domain audio adaptive denoising method, by extracting MDCT coefficient from MP3 audio frequency, carry out prior modeling to the distribution of MDCT coefficient, and construct estimator, realize the MP3 audio frequency that contains noise Noise reduction processing.

The technical scheme adopted by the present invention to solve the technical problem is as follows: first extract MDCT coefficients from MP3 audio data, then perform prior probability modeling on the MDCT coefficients, and then construct a noise attenuation estimator. At the same time, it detects the silent segment of the MP3 audio, and adjusts the degree of noise attenuation for the noise-containing frequency band according to the attenuation proportion of the silent segment.

The technical solution adopted by the present invention to solve the technical problem can be further improved. First extract the MDCT coefficients from the MP3 audio data, and then analyze the characteristics of the MDCT coefficients. According to the characteristics of the MDCT coefficients, the normal anti-Gaussian distribution function suitable for modeling the prior probability distribution of the MDCT coefficient distribution is selected, and then according to the Bayesian maximum posterior The experimental probability theory is used to construct the noise attenuation estimator. At the same time, the MDCT spectral energy feature is used to detect the silent segment of the MP3 audio, and the degree of noise attenuation during the noise reduction process is adjusted according to the attenuation proportion of the silent segment. The method specifically includes the following steps:

1), the preprocessing of MP3 compressed audio containing noise, including four parts: decoding the MP3 frame header, obtaining side information, obtaining main data and scaling factor, Huffman decoding and dequantization;

2), extract the MDCT coefficients, and perform amplitude mapping processing: find out the MDCT coefficients of each frame with two granularities from the dequantized MP3 frame, average the MDCT coefficients of the two particles according to the frequency points, and construct each frame The MDCT spectral coefficient of the audio, and the amplitude range of the MDCT coefficient is mapped to between 0-L;

3) Perform prior modeling on the distribution of MDCT coefficients and construct a maximum a posteriori probability estimator: respectively analyze the distribution of MDCT coefficients without noise and MDCT coefficients with noise, and obtain the MDCT coefficients without noise statistical properties. According to the sparse statistical properties of MDCT coefficients, the normal inverse Gaussian (NIG) distribution function is used to model the MDCT coefficients a priori. According to the Bayesian maximum a posteriori probability criterion, an estimator based on the NIG prior distribution model is derived.

4), silent segment detection: extract the spectral energy feature based on MDCT coefficients, and detect the silent segment in the MP3 audio according to the energy feature parameter;

5), adaptive iterative estimation: use the estimator in 3) to estimate the MP3 containing noise, and adjust the number of iterative estimation adaptively through the attenuation factor of the silent segment detected in 4).

The beneficial effects of the present invention are: directly based on the MP3 compressed domain, MP3 audio is subjected to noise reduction processing, and compared with the traditional method of decoding MP3 compressed audio into non-compressed wave audio and then performing noise reduction processing, the method proposed by the present invention is more effective. It is simple and saves calculation time; study the distribution characteristics of MDCT coefficients of MP3 audio, and select a function suitable for prior modeling of the distribution of MDCT coefficients. Experimental results show that the selected function can effectively fit the distribution of MDCT coefficients; Moreover, the noise attenuation estimator designed based on the prior probability distribution function of MDCT coefficients can effectively realize the noise reduction of MP3 compressed audio; at the same time, the silent segment in MP3 audio is detected by using the MDCT spectral energy feature, and then the attenuation factor of the silent segment Adaptive control of the degree of noise attenuation can not only effectively solve the problem of introducing audio noise caused by over-attenuation or under-attenuation in the noise reduction process, but also the audio after noise reduction has a good effect.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Detailed ways

A preferred embodiment of a kind of MP3 compressed domain audio adaptive noise reduction method of the present invention is described as follows in conjunction with accompanying drawing: A kind of MP3 compressed domain audio adaptive noise reduction method is divided into five steps altogether:

Step 1: Preprocessing of noisy MP3 compressed audio

Preprocessing of noisy MP3 compressed audio includes four parts: decoding MP3 frame header, obtaining side information, reading main data and scaling factor, Huffman decoding and dequantization.

1. Acquisition of synchronous data stream and frame header information

A), according to the MP3 encoding format, search for synchronous information from the MP3 data stream;

B), according to synchronous information, find the starting position of each frame data in the MP3 data flow;

C), after determining the starting position of the data frame, obtain the frame header information Head;

2. Acquisition of side information

A), according to the encoding format of MP3 frame header, determine the starting position of edge information in MP3 frame header;

B), obtain the side information Side from the MP3 frame header information Head;

3. Reading of MP3 main data and scaling factor

A), calculate the length L of the main data according to the side information Side;

B), according to the offset of the main data in the frame header information Head, determine the starting position of the MP3 main data;

C), obtaining the main data D with a total length of L from the current frame;

D), extract the scaling factor Scale from the master data D;

4. Huffman decoding and dequantization

A), determine the start and end positions of the Huffman decoded data according to the side information Side;

B), carry out Huffman decoding to MP3 main data D, obtain the Huffman decoding result F [32,18] of 32*18 dimensions;

C) Dequantize the data in the Huffman decoding result F[32, 18].

The second step: MDCT coefficient extraction and amplitude mapping processing

1. Construct the modified discrete cosine transform MDCT coefficients of each frame of audio

A), allocate the n*576 size storage space MDCT ₀ [n, 576], MDCT ₁ [n, 576] for depositing the MDCT coefficient of two granularities of a frame of MP3 audio frequency, wherein n is the frame number of MP3 audio frequency ;

B), from the array F, find the MDCT coefficients of the two granularities of the same frame of audio, rearrange according to the principle of frequency from low to high, and obtain MDCT ₀ [i, j], MDCT ₁ [i, j];

C), calculate the average value of the MDCT coefficients at two granularity identical frequency points in the same frame of audio, as the MDCT coefficient value M[i, j] of this frame of audio;

$\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; MDCT</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; MDCT</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, MDCT ₀ [i, j] and MDCT ₁ [i, j] are the 0th granularity and the jth MDCT spectrum value of the first granularity of the i-th audio frame respectively. M[i, j] is the j-th average MDCT spectral value of the i-th frame of audio.

2. MDCT coefficient amplitude range mapping

Linearly map the amplitude of MDCT coefficients in the range of 0-1 to 0-P, which is convenient for studying the statistical distribution of MDCT coefficients and the corresponding fitting functions

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; P</span>}$

In the formula, x′ _ij is the j-th MDCT spectral value of the i-th frame audio after amplitude mapping, M[i, j] is the j-th average MDCT spectral value of the i-th frame audio obtained in 1, M _min is the smallest MDCT spectral coefficient, M _max is the largest MDCT spectral coefficient, and P is the maximum magnitude after mapping.

Step 3: Prior Modeling of MDCT Coefficients and Maximum Posteriori Estimator

1. Analyze the distribution characteristics of MDCT

2. Calculate the probability distribution function of the MDCT coefficient

After the analysis of 1 shows that the distribution of MDCT has a sparse characteristic, the normal inverse Gaussian distribution function is used to simulate the distribution of MDCT coefficients, and the MDCT probability distribution function is expressed as:

$\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;\&delta;</span>}}{\mathrm{<span\; class="notranslate">\; \&pi;q</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; \&alpha;q</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>$

K_λ(·)是索引为λ的第二阶修正贝塞尔函数，K₁(·)是索引为1的第二阶修正贝塞尔函数，

0≤|β|＜α，δ＞0，-∞＜μ＜∞。其中，α为衰减因子，δ为尺度因子，μ为均值，β为倾斜因子。In the formula,

K _λ (·) is the second-order modified Bessel function whose index is λ, K ₁ (·) is the second-order modified Bessel function whose index is 1,

0≤|β|<α, δ>0, -∞<μ<∞. Among them, α is the attenuation factor, δ is the scale factor, μ is the mean value, and β is the tilt factor.

3. Analyze the influence of parameters [α, δ, β, μ] ^T on the characteristics of normal inverse Gaussian distribution

4. Parameter estimation

To use the normal inverse Gaussian distribution function in 2 to fit the probability distribution of MDCT coefficients, the parameters [α, δ, β, μ] ^T need to be estimated.

均值μ，倾斜因子β假设加入的噪声为零均值高斯白噪声，含噪音频的前几帧为纯噪声帧，由纯噪声帧估计噪声MDCT系数的方差

并对含有噪声的MDCT系数计算均值μ。MP3音频信号的MDCT系数呈对称分布，因此，假设倾斜因子β＝0。A), calculate the variance

Mean value μ, slope factor β assume that the noise added is zero-mean Gaussian white noise, and the first few frames of the noise-containing frequency are pure noise frames, and the variance of the noise MDCT coefficients is estimated from the pure noise frames

And the mean value μ is calculated for the MDCT coefficients containing noise. The MDCT coefficients of the MP3 audio signal are distributed symmetrically, therefore, it is assumed that the tilt factor β=0.

B), calculate the attenuation factor α, scale factor δ

峭度系数为其中

相应的衰减因子α、尺度因子δ可通过如下式子进行估计：The skew coefficient of the NIG distribution model of the noise-free MDCT coefficients is

The kurtosis coefficient is in

The corresponding attenuation factor α and scale factor δ can be estimated by the following formula:

$\mathrm{<span\; class="notranslate">\; \&delta;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&times;</span>\sqrt{{\mathrm{<span\; class="notranslate">\; \&gamma;\&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>}$

$\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\sqrt{{\mathrm{<span\; class="notranslate">\; \&gamma;\&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; /</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; k</span>}}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}$

C₁，C₂为用于控制衰减因子α、尺度因子δ的幅值，使NIG能有效地拟合MDCT系数的分布。in, are the 2nd to 4th order cumulants of the noisy MDCT coefficients, respectively,

C ₁ and C ₂ are used to control the amplitude of attenuation factor α and scale factor δ, so that NIG can effectively fit the distribution of MDCT coefficients.

C), estimated parameters C ₁ and C ₂

For different audio types and under different signal-to-noise ratio conditions, the simulation errors of different values of C ₁ and C ₂ on the distribution of MDCT coefficients are counted, and finally the best values C ₁ = 0.1, C ₂ = 0.1 are obtained; therefore, there is an attenuation factor α , the estimation formula of scale factor δ is:

$\mathrm{<span\; class="notranslate">\; \&delta;</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0.1</span>}<span\; class="notranslate">\; \&times;</span>\sqrt{{\mathrm{<span\; class="notranslate">\; \&gamma;\&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>}$

$\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0.1</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\sqrt{{\mathrm{<span\; class="notranslate">\; \&gamma;\&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; /</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; k</span>}}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}$

5. Design an attenuation estimator

According to the Bayesian maximum a posteriori probability criterion, the estimation function based on the NIG prior distribution model is designed:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; \&zeta;</span>}}<span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; \&zeta;</span>}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}$

K_λ(·)是索引为λ的第二阶修正贝塞尔函数，

为对含有噪声的MP3音频数据y进行衰减得到的无噪MP3音频数据。In the formula,

K _λ ( ) is the second-order modified Bessel function with index λ,

It is the noise-free MP3 audio data obtained by attenuating the MP3 audio data y containing noise.

Correspondingly, the attenuation factor of MP3 audio containing noise can be obtained as:

$\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; \&zeta;</span>}}$

Step 4: Silent segment detection

1. Extraction of MDCT spectral features

The spectral energy of the MDCT coefficients is calculated as follows:

$\mathrm{<span\; class="notranslate">\; EM</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>$

Wherein, EM (i) is the MDCT spectral energy of the i-th frame audio, M (i, j) is the j MDCT spectral mean value of the i-th frame audio, and N is the number of points N=576 of the MDCT coefficient of a frame of audio. For the whole MP3 audio segment, the MDCT spectrum energy of each frame of the audio segment forms a corresponding feature vector EM=[EM (0), EM (1), ... EM (N-1)], that is, EM is the MDCT of the audio segment Spectral energy envelope.

2. Adjustment of judgment threshold

A), initialize the decision threshold, and use the mean value of the MDCT spectrum energy envelope of the entire signal as the initial decision threshold L _th

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; the\; th</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; EM</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

In the formula, EM(i) is the MDCT spectrum energy of the i-th audio frame, N is the frame number of the audio segment, and L _th is the initial decision threshold.

B), threshold adjustment: in the MDCT spectrum envelope EM of the audio segment, all frames less than the decision threshold L _th are processed as noise frames, with

EM _noise (i) = EM (i) if EM (i) < I _th

In the formula, EM _noise (i) indicates that the MDCT spectral energy value of the i-th frame audio is the MDCT spectral energy value of the noise frame.

Initialize the mean and mean square error of the noise spectrum sequence, denoted as L _noise and S _noise respectively,

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; noise</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; Em</span>}}_{\mathrm{<span\; class="notranslate">\; noise</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; noise</span>}}<span\; class="notranslate">\; =</span>\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; EM</span>}}_{\mathrm{<span\; class="notranslate">\; noise</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; noise</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

In the formula, EM _noise (i) represents the MDCT spectrum energy value of the i-th noise frame, L _noise and S _noise are the mean value and mean square error of the noise energy sequence respectively, and M is the number of frames of the noise segment.

On the basis of obtaining the mean value L _noise and the mean square error S _noise of the energy sequence of the noise frame, the decision threshold L _th is readjusted.

L _th ＝C ₀ ×(L _noise +C ₁ ×S _noise )

Among them, C0 and C1 are empirical constants. In the experiment, C0=1.001, and the value of C1 is adjusted between 1.5 and 2.0. After the decision threshold L _th is adjusted, the noise and speech frames are re-distinguished, and the mean value L _noise and the mean square error S _noise of the noise spectrum energy sequence are recalculated, and then the decision threshold is adjusted. Repeat this until the decision threshold is stable.

3. Fusion of active endpoints

A), judge the mute frame/non-silent frame according to the threshold

${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; type</span>}}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; EM</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; <</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; the\; th</span>}}\\ \mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& \mathrm{<span\; class="notranslate">\; EM</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&Greater\; Equal;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; the\; th</span>}}\end{array}\right.$

Among them, E _type [i] is the type of the i-th frame audio, and EM[i] is the MDCT spectrum energy value of the i-th frame audio; the value of the audio type E _type [i] is 0 to indicate a silent frame, and the type E _type [i] A value of 1 indicates an active audio frame.

B), calculating the number of frames F _N included in the silent segment;

C), if F _N <10, this segment is a pause between continuous active audio segments, and is merged into the corresponding audio segment;

Step 5: Adaptive iterative decay

1. Calculate the attenuation value of the silent segment for the silent segment detected in the fourth step by the attenuation function obtained in the third step;

2. Calculate the average attenuation value of the silent segment in 1

3. Use the attenuation function obtained in the third step to attenuate the MDCT coefficients of the noisy MP3 audio;

自适应调整迭代估计的次数：重复步骤1、2、3，当满足下面条件，则停止迭代，降噪完成：4. From the average attenuation value of the silent segment

Adaptively adjust the number of iterative estimates: Repeat steps 1, 2, and 3. When the following conditions are met, the iteration is stopped and the noise reduction is completed:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; c</span>}$

为静音段的平均衰减值，a_min为整段音频的最小衰减因子，可以通过MDCT系数的高频段获得。C用于控制余留分量，取C＝0.001，见附图1。in,

is the average attenuation value of the silent segment, and a _min is the minimum attenuation factor of the entire audio segment, which can be obtained through the high frequency band of the MDCT coefficient. C is used to control the residual component, and C=0.001, see Figure 1.

Experimental results

In this experiment, the audio data broadcast by CCTV was used for the experiment. The format of the audio data is MP3, and the sampling frequency is 44.1KHz. The audio types are: speech, music, audio mixed with speech and music. Choose 20 songs for each type of audio. Add different degrees of Gaussian white noise to various types of audio, and use the adaptive noise reduction algorithm proposed in this study to process the noisy MP3 audio. The calculation method of the signal-to-noise ratio SNR after noise reduction processing is:

$\mathrm{<span\; class="notranslate">\; SNR</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}\frac{\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Wherein, x(n) is the PCM data obtained by decoding MP3 audio without noise, PCM data decoded for noise-reduced MP3 audio. The SNR comparison results before and after noise reduction processing are shown in Table 1:

Table 1: SNR comparison of MP3 audio before and after noise reduction

MP3 audio signal SNR before noise reduction SNR after noise reduction Average SNR Gain Music 1 -5db 8.11db 13.11db Music 2 0db 11.40db 11.40db Music 3 5db 14.89db 9.89db Music 4 10db 17.93db 7.93db Music 5 15db 22.57db 7.57db Voice 1 -5db 8.12db 13.12db Voice 2 0db 10.78db 10.78db Music + Voice 1 -5db 6.26db 11.26db Music + Voice 2 0db 9.13db 9.13db

A large number of statistical experiments show that the noise reduction method based on the MP3 compressed domain audio of the present invention can be directly based on the MP3 compressed domain, and effectively implement noise reduction processing on different types of MP3 audio containing noise. The signal-to-noise ratio of the MP3 audio after the noise reduction processing is greatly improved, and the processed audio has a good auditory perception effect. This study solves the problem of noise reduction directly based on MP3 compressed domain audio, and also proposes a new idea for the research of anti-noise algorithm for MP3 audio classification retrieval.

Claims (5)

1. An MP3 compressed domain audio adaptive noise reduction method is characterized in that: firstly, extracting MDCT coefficients representing the frequency domain characteristics of original audio from MP3 compressed audio, then analyzing the sparse statistical characteristics of the MDCT coefficients, carrying out prior modeling on the MDCT coefficients by adopting a normal inverse Gaussian NIG distribution function, and designing a maximum posterior probability estimator based on the normal inverse Gaussian NIG prior probability model by utilizing a Bayesian criterion to obtain attenuation factors of corresponding audio segments; finally, in a noise reduction part, detecting a mute section in the MP3 audio by using the MDCT spectrum energy characteristics, and adaptively controlling the iteration times of noise attenuation through the attenuation weight of the detected mute section, thereby realizing the adaptive noise reduction of the MP3 compressed audio;

the specific operation steps are as follows:

1) preprocessing the MP3 compressed audio containing noise, including decoding the MP3 frame header, acquiring side information, acquiring main data and a scaling factor, performing Huffman decoding and performing inverse quantization;

2) extracting MDCT coefficients and carrying out amplitude mapping: the MDCT coefficients of two granularities of each frame are found out from the MP3 frame after inverse quantization, the MDCT coefficients of two granules are averaged according to frequency points, the MDCT spectrum coefficient of each frame of audio is constructed, and the amplitude range of the MDCT coefficient is mapped between 0 and L;

3) carrying out prior modeling on the distribution of the MDCT coefficient and constructing a maximum posterior probability estimator: respectively analyzing the distribution conditions of the MDCT coefficient without noise and the MDCT coefficient with noise to obtain the statistical characteristics of the MDCT coefficient without noise; according to the sparse statistical characteristics of the MDCT coefficient, carrying out prior modeling on the MDCT coefficient by utilizing a normal inverse Gaussian NIG distribution function; designing a noise attenuation estimator based on an NIG prior distribution model according to a Bayes maximum posterior probability criterion;

4) and mute section detection: extracting the spectral energy characteristics based on the MDCT coefficient, and detecting a mute section in the MP3 audio according to the MDCT spectral energy characteristic parameters;

5) and self-adaptive iterative estimation: estimating the MP3 containing noise by using the estimator in the step 3), and adaptively adjusting the number of times of iterative estimation by the attenuation factor of the mute section detected in the step 4);

the specific steps of carrying out prior modeling on the distribution of the MDCT coefficients in the step 3) and constructing the maximum posterior probability estimator are as follows:

firstly, analyzing the distribution characteristics of MDCT;

secondly, calculating a probability distribution function of the MDCT coefficient;

after the MDCT distribution obtained through the analysis of the step I has the sparse characteristic, the distribution of the MDCT coefficient is simulated by adopting a normal inverse Gaussian distribution function, and the MDCT probability distribution function is obtained and expressed as:

<math> <mrow> <mi>p</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mi>&alpha;&delta;</mi> <mrow> <mi>&pi;q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mi>exp</mi> <mo>[</mo> <mi>h</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>]</mo> <msub> <mi>K</mi> <mn>1</mn> </msub> <mo>[</mo> <mi>&alpha;q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>]</mo> </mrow> </math>

in the formula:

K_λ(. is a second order modified Bessel function with an index of λ, K₁(. cndot.) is a second order modified Bezier function with an index of 1, <math> <mrow> <mi>h</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mi>&delta;</mi> <msqrt> <msup> <mi>&alpha;</mi> <mn>2</mn> </msup> <mo>-</mo> <msup> <mi>&beta;</mi> <mn>2</mn> </msup> </msqrt> <mo>+</mo> <mi>&beta;</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>-</mo> <mi>&mu;</mi> <mo>)</mo> </mrow> <mo>,</mo> </mrow> </math> <math> <mrow> <mi>q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <msup> <mi>&delta;</mi> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <mi>x</mi> <mo>-</mo> <mi>&mu;</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> <mo>,</mo> </mrow> </math> beta is more than or equal to 0 and less than alpha, delta is more than 0, and mu is more than infinity; wherein alpha is an attenuation factor, delta is a scale factor, mu is a mean value, and beta is a gradient factor;

analyzing the influence of four parameters of attenuation factor alpha, scale factor delta, mean value mu and inclination factor beta on the normal inverse Gaussian distribution characteristic;

fourthly, parameter estimation

The normal inverse Gaussian distribution function in the step II is adopted to fit the probability distribution of the MDCT coefficient, and four parameters [ alpha, delta, beta, mu ] for determining the shape of the normal inverse Gaussian distribution are required]^TAnd (3) estimating:

A) calculating the variance

Mean value μ, skew factor β

Assuming that the added noise is white Gaussian noise with zero mean, the first frames of the noisy audio are pure noise frames, and the variance of MDCT (modified discrete cosine transform) coefficients of the noise is estimated from the pure noise framesAnd calculating the average value mu of the MDCT coefficients containing noise, wherein the MDCT coefficients of the MP3 audio signal are distributed symmetrically, so that the inclination factor beta is assumed to be 0;

B) calculating attenuation factor alpha and scale factor delta

Estimating an attenuation factor alpha and a scale factor delta by adopting a skew coefficient and a kurtosis coefficient of an NIG distribution model; skew coefficients of the NIG distribution model of noise-free MDCT coefficients are

Kurtosis coefficient of

Wherein

Corresponding attenuation factor alpha and scale factor delta-passThe estimation is performed by the following equation:

<math> <mrow> <mi>&delta;</mi> <mo>=</mo> <msub> <mi>C</mi> <mn>1</mn> </msub> <mo>&times;</mo> <msqrt> <mi>&gamma;</mi> <msub> <mi>&beta;</mi> <mn>2</mn> </msub> <mo>|</mo> <mn>1</mn> <mo>-</mo> <msup> <mi>&eta;</mi> <mn>2</mn> </msup> <mo>|</mo> </msqrt> </mrow> </math>

<math> <mrow> <mi>&alpha;</mi> <mo>=</mo> <msub> <mi>C</mi> <mn>2</mn> </msub> <mo>&times;</mo> <msubsup> <mi>&beta;</mi> <mn>2</mn> <mn>2</mn> </msubsup> <msqrt> <mi>&gamma;</mi> <msub> <mi>&beta;</mi> <mn>2</mn> </msub> </msqrt> <mo>/</mo> <msub> <mover> <mi>k</mi> <mrow> <mo>^</mo> <mi></mi> </mrow> </mover> <mn>4</mn> </msub> </mrow> </math>

wherein,respectively 2 to 4 order cumulants of MDCT coefficients containing noise,

parameter C₁，C₂For weighting the attenuation factor alpha and the scale factor delta, selecting proper C₁，C₂Values that enable the NIG to effectively fit the distribution of MDCT coefficients;

C) estimating attenuation factor and weight C of scale factor₁、C₂

For different audio types and under different signal-to-noise ratios, counting C₁、C₂The fitting error of MDCT coefficient distribution is obtained when different values are taken, and finally the optimal value C is obtained₁＝0.1，C₂0.1; therefore, the estimation formula of the attenuation factor alpha and the scale factor delta is as follows:

<math> <mrow> <mi>&delta;</mi> <mo>=</mo> <mn>0.1</mn> <mo>&times;</mo> <msqrt> <mi>&gamma;</mi> <msub> <mi>&beta;</mi> <mn>2</mn> </msub> <mo>|</mo> <mn>1</mn> <mo>-</mo> <msup> <mi>&eta;</mi> <mn>2</mn> </msup> <mo>|</mo> </msqrt> </mrow> </math>

<math> <mrow> <mi>&alpha;</mi> <mo>=</mo> <mn>0.1</mn> <mo>&times;</mo> <msubsup> <mi>&beta;</mi> <mn>2</mn> <mn>2</mn> </msubsup> <msqrt> <mi>&gamma;</mi> <msub> <mi>&beta;</mi> <mn>2</mn> </msub> </msqrt> <mo>/</mo> <msub> <mover> <mi>k</mi> <mo>^</mo> </mover> <mn>4</mn> </msub> </mrow> </math>

designing a noise attenuation function based on an NIG prior distribution model according to a Bayes maximum posterior probability criterion:

<math> <mrow> <mover> <mi>x</mi> <mo>^</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>1</mn> <mo>+</mo> <msubsup> <mi>&sigma;</mi> <mi>&epsiv;</mi> <mn>2</mn> </msubsup> <mi>&zeta;</mi> </mrow> </mfrac> <mo>&times;</mo> <mrow> <mo>(</mo> <mi>y</mi> <mo>+</mo> <msubsup> <mi>&sigma;</mi> <mi>&epsiv;</mi> <mn>2</mn> </msubsup> <mi>&beta;</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>1</mn> <mo>+</mo> <msubsup> <mi>&sigma;</mi> <mi>&epsiv;</mi> <mn>2</mn> </msubsup> <mi>&zeta;</mi> </mrow> </mfrac> <mo>&times;</mo> <mi>y</mi> </mrow> </math>

in the formula, <math> <mrow> <mi>&zeta;</mi> <mo>=</mo> <mfrac> <mi>&alpha;</mi> <mrow> <mi>q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>&times;</mo> <mfrac> <mrow> <msub> <mi>K</mi> <mn>0</mn> </msub> <mrow> <mo>(</mo> <mi>&alpha;q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>K</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <mi>&alpha;q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> </mrow> <mrow> <mn>2</mn> <msub> <mi>K</mi> <mn>1</mn> </msub> <mrow> <mo>(</mo> <mi>&alpha;q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>+</mo> <mfrac> <mn>1</mn> <msup> <mrow> <mo>[</mo> <mi>q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>]</mo> </mrow> <mn>2</mn> </msup> </mfrac> <mo>,</mo> </mrow> </math> <math> <mrow> <mi>q</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <msup> <mi>&delta;</mi> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <mi>x</mi> <mo>-</mo> <mi>&mu;</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> <mo>,</mo> </mrow> </math> k λ (·) is a second-order modified bessel function with index λ,

noise-free MP3 audio data obtained by attenuating the MP3 audio data y containing noise;

accordingly, the attenuation factor for MP3 audio with noise can be obtained as:

<math> <mrow> <mi>a</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>1</mn> <mo>+</mo> <msubsup> <mi>&sigma;</mi> <mi>&epsiv;</mi> <mn>2</mn> </msubsup> <mi>&zeta;</mi> </mrow> </mfrac> <mo>.</mo> </mrow> </math>

2. the MP3 compressed domain audio adaptive noise reduction method of claim 1, wherein: the specific steps of performing MP3 compressed audio preprocessing in step 1) are as follows:

firstly, acquiring synchronous data stream and frame header information;

A) searching synchronous information from MP3 data stream according to MP3 encoding format;

B) finding the initial position of each frame data in the MP3 data stream according to the synchronization information;

C) after determining the initial position of the data frame, acquiring frame header information Head;

secondly, acquiring side information from the frame header information obtained by decoding

A) Determining the initial position of the side information in the MP3 frame header according to the encoding format of the MP3 frame header;

B) acquiring Side information Side from MP3 frame header information Head;

extracting MP3 main data and scaling factor

A) Calculating the length L of the main data according to the Side information Side;

B) determining the initial position of the MP3 main data according to the offset of the main data in the header information Head;

C) acquiring main data D with the total length L from the current frame;

D) extracting a scaling factor Scale from the main data D;

fourthly, carrying out Huffman decoding and inverse quantization on the MP3 main data stream

A) Determining the starting position and the ending position of the Huffman decoding data according to the Side information Side;

B) performing Huffman decoding on the MP3 main data D to obtain a 32-by-18-dimensional Huffman decoding result F [32, 18 ];

C) and inverse quantizing the data in the Huffman decoding result F [32, 18 ].

3. The MP3 compressed domain audio adaptive noise reduction method of claim 2, wherein: the MDCT coefficient extraction and amplitude mapping process in step 2) specifically includes the following steps:

firstly, constructing Modified Discrete Cosine Transform (MDCT) coefficient of each frame of audio

A) Memory space MDCT of size n x 576 for storing MDCT coefficients of two granularities of one frame MP3 audio₀[n，576]，MDCT₁[n，576]Wherein n is the number of frames of MP3 audio;

B) respectively finding MDCT coefficients of two granularities of the same frame of audio from the array F, rearranging according to the principle that the frequency is from low to high to obtain the MDCT₀[i，j]，MDCT₁[i，j]；

C) Calculating the average value of MDCT coefficients at two frequency points with the same granularity in the same frame audio to be used as an MDCT coefficient value M [ i, j ] of the frame audio;

$M[i,j]=\frac{{\mathrm{MDCT}}_0[i,j]+{\mathrm{MDCT}}_1[i,j]}{2}$

wherein the MDCT₀[i，j]，MDCT₁[i，j]The jth MDCT spectrum values of the 0 th granularity and the 1 st granularity of the ith frame of audio respectively; m [ i, j ]]The j average MDCT spectrum value of the i frame audio;

and secondly, mapping the amplitude range of the MDCT coefficient: the amplitude of the MDCT coefficient is linearly mapped between 0 and P in the range of 0 to 1, so that the statistical distribution of the MDCT coefficient and the corresponding fitting function are conveniently researched:

<math> <mrow> <msubsup> <mi>x</mi> <mi>ij</mi> <mo>&prime;</mo> </msubsup> <mo>=</mo> <mfrac> <mrow> <mi>M</mi> <mo>[</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>]</mo> <mo>-</mo> <msub> <mi>M</mi> <mi>min</mi> </msub> </mrow> <mrow> <msub> <mi>M</mi> <mi>max</mi> </msub> <mo>-</mo> <msub> <mi>M</mi> <mi>min</mi> </msub> </mrow> </mfrac> <mo>&times;</mo> <mi>P</mi> </mrow> </math>

x 'in the formula'_ijFor the j (th) MDCT spectrum value of the i (th) frame audio after amplitude mapping, M [ i, j]Is the j average MDCT spectrum value of the i frame audio obtained by (r)_minIs the smallest MDCT spectral coefficient, M_maxIs the largest MDCT spectral coefficient, and P is the maximum amplitude after mapping.

4. The MP3 compressed domain audio adaptive noise reduction method of claim 1, wherein: the step 4) of detecting the mute section comprises the following specific steps:

extracting spectral features based on MDCT coefficients

<math> <mrow> <mi>EM</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>j</mi> <mo>=</mo> <mn>0</mn> </mrow> <mrow> <mi>N</mi> <mo>-</mo> <mn>1</mn> </mrow> </munderover> <msup> <mi>M</mi> <mn>2</mn> </msup> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> </mrow> </math>

Wherein EM (i) is the MDCT spectral energy of the i-th frame of audio, M (i, j) is the jth MDCT spectral mean of the i-th frame of audio, N is the number N of MDCT coefficients of one frame of audio 576, and for the entire MP3 audio segment, the MDCT spectral energy of each frame of the audio segment constitutes a corresponding feature vector EM [ EM (0), EM (1),.. EM (N-1) ], i.e. EM is the MDCT spectral energy envelope of the audio segment;

secondly, adjusting a decision threshold according to the MDCT spectrum energy characteristics

A) Initializing a decision threshold, and taking the average value of the MDCT spectrum energy envelope of the whole signal as an initial decision threshold L_th

<math> <mrow> <msub> <mi>L</mi> <mi>th</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>0</mn> </mrow> <mrow> <mi>N</mi> <mo>-</mo> <mn>1</mn> </mrow> </munderover> <mi>EM</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </mrow> </math>

Where EM (i) is the MDCT spectral energy of the i frame of audio, N represents the number of frames of the audio segment, L_thIs an initial decision threshold;

B) and adjusting a threshold: all MDCT spectrum envelopes EM of the audio segment are smaller than a judgment threshold L_thThe frame is processed as a noise frame, having

EM_noise(i)＝EM(i) if EM(i)＜L_th

In the formula, EM_noise(i) The MDCT spectral energy value representing the i-th frame audio is the MDCT spectral energy value of the noise frame,

mean and mean square error of the initialized noise spectrum sequence, respectively denoted as L_noiseAnd S_noise，

<math> <mrow> <msub> <mi>L</mi> <mi>noise</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mi>M</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>0</mn> </mrow> <mrow> <mi>M</mi> <mo>-</mo> <mn>1</mn> </mrow> </munderover> <msub> <mi>EM</mi> <mi>noise</mi> </msub> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </mrow> </math>

<math> <mrow> <msub> <mi>S</mi> <mi>noise</mi> </msub> <mo>=</mo> <msqrt> <mfrac> <mn>1</mn> <mi>M</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>0</mn> </mrow> <mrow> <mi>M</mi> <mo>-</mo> <mn>1</mn> </mrow> </munderover> <msup> <mrow> <mo>(</mo> <msub> <mi>EM</mi> <mi>noise</mi> </msub> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>L</mi> <mi>noise</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mrow> </math>

In the formula, EM_noise(i) MDCT spectral energy value, L, representing the i-th noise frame_noise、S_noiseRespectively representing the mean value and the mean square error of the noise energy sequence, wherein M is the frame number of the noise section;

obtaining the mean value L of the energy sequence of the noise frame_noiseSum mean square error S_noiseOn the basis, readjust the decision threshold L_th，

L_th＝C₀×(L_noise+C₁×S_noise)

Wherein C0 and C1 are empirical constants, C0 is 1.001, and C1 is adjusted between 1.5 and 2.0; adjusted decision threshold value L_thThen, the noise and the voice frame are distinguished again, and the mean value L of the noise spectrum energy sequence is calculated again_noiseSum mean square error S_noiseThen adjusting a judgment threshold value; repeating the above steps until the decision threshold is stable;

③ fusion of active endpoints

A) Judging the mute frame/non-mute frame according to the threshold

<math> <mrow> <msub> <mi>E</mi> <mi>type</mi> </msub> <mo>[</mo> <mi>i</mi> <mo>]</mo> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <mn>0</mn> <mo>,</mo> </mtd> <mtd> <mi>EM</mi> <mo>[</mo> <mi>i</mi> <mo>]</mo> <mo>&lt;</mo> <msub> <mi>L</mi> <mi>th</mi> </msub> </mtd> </mtr> <mtr> <mtd> <mn>1</mn> <mo>,</mo> </mtd> <mtd> <mi>EM</mi> <mo>[</mo> <mi>i</mi> <mo>]</mo> <mo>&GreaterEqual;</mo> <msub> <mi>L</mi> <mi>th</mi> </msub> </mtd> </mtr> </mtable> </mfenced> </mrow> </math>

E_type[i]Type of audio for frame i, EM [ i ]]MDCT spectrum energy value of the ith frame audio; audio type E_type[i]A value of 0 indicating a silent frame, type E_type[i]A value of 1 indicates an active audio frame;

B) calculating the number of frames F contained in the mute section_N；

C) If F_N< 10, which is a pause between consecutive active audio segments, incorporated into the corresponding audio segment.

5. The MP3 compressed domain audio adaptive noise reduction method of claim 1, wherein: the step 5) of the adaptive iterative estimation specifically comprises the following steps:

calculating the attenuation value of the mute section detected in the step 4) by the attenuation function obtained in the step (3);

secondly, calculating the average attenuation value of the mute section in the step I

Using the attenuation function obtained in the step 3) to attenuate the MDCT coefficient of the MP3 audio containing noise;

fourthly, average attenuation value of silence segmentAdaptively adjusting the number of iterative estimation: repeating the first step, the second step and the third step, stopping iteration and finishing noise reduction when the following conditions are met:

<math> <mrow> <mover> <mi>a</mi> <mo>&OverBar;</mo> </mover> <mo>&le;</mo> <msub> <mi>a</mi> <mi>min</mi> </msub> <mo>+</mo> <mi>c</mi> </mrow> </math>

is the average attenuation value of the silence segment, a_minThe minimum attenuation factor for the whole audio can be obtained by a high-frequency segment of MDCT coefficients, C is used to control the residual component, and C is 0.001.

Images