CN104021791B - Detecting method based on digital audio waveform sudden changes - Google Patents

Abstract

The invention discloses a detection method for sudden changes in audio waveforms, which is a statistical discrimination method proposed according to the characteristics of ridges in speech spectrum caused by sudden changes in audio waveforms, and belongs to the field of multimedia information security. The method of the present invention aims at the waveform mutation detection of digital audio due to the copy and paste operation, analyzes the ridge-peak span changes before and after the audio splicing point in the logarithmic domain of the spectrum, and constructs a ridge-peak factor to describe the ridge-peak bandwidth of a frame of audio logarithmic spectrum to represent Short-term energy change characteristics, and use the difference operator to distinguish the sudden sound of the audio from the ridge factor identification caused by the sudden change of the audio waveform. The method of the present invention comprises the following steps: performing short-time Fourier transform and logarithmic transformation on the audio signal to obtain the audio spectrum in the logarithmic domain; calculating the ridge factor of each frame in the spectrum; performing differential transformation on the ridge factor The results are tested and judged. The invention can effectively identify sudden changes in audio waveforms, and provides an effective method for detecting the boundaries of digital audio editing operations.

Description

Detection method of digital audio waveform mutation

technical field

The invention relates to the field of multimedia information security, in particular to a method for detecting sudden changes in digital audio waveforms.

Background technique

The digitalization of multimedia technology and the progress of transmission technology make the application of digital audio grow rapidly. Digital audio is easy to transmit and disseminate through physical or electronic systems, but at the same time these advantages also bring many new problems: digital audio may suffer from intentional or unintentional tampering in the process of recording and copying. Whether it is intentional tampering that destroys integrity and authenticity, or errors that may occur during transmission and dissemination, it will damage the originality of its own information data, especially in some information with special significance, such as court evidence, department Once important content such as confidential documents and historical document backups are maliciously tampered with, serious consequences will result.

Digital voice copy-paste/delete tampering is to copy a segment of voice in digital voice to another segment of voice, or to delete a segment in a segment of voice. It is a simple and effective method of tampering to change important information of voice, as shown in the attached picture of the manual 1, 2 shown. Because there are consistent or similar noises and voiceprints in the same speech, it is difficult for the human ear to distinguish the falsified speech segments. Detecting this form of tampering has important practical significance for judging the authenticity of speech.

Contents of the invention

The technical problem to be solved by the present invention is to provide a detection method for a digital audio waveform mutation, which can judge whether a given section of speech has undergone a audio waveform mutation caused by copying and pasting a section of speech or deleting a section of speech. The time range of voice tampering is located, so as to confirm the authenticity of the voice, so as to overcome the shortcomings of the existing technology.

The present invention is realized in the following way: the detection method of digital audio waveform mutation includes the following steps: 1) Transform the audio signal to obtain the audio spectrum Y in the logarithmic domain, perform logarithmic transformation on the obtained audio spectrum, and obtain the logarithmic spectrum Spectrum G ; 2) Calculate the energy binarization of the logarithmic spectrum G; 3) Calculate the audio mutation coefficient σ _t of each frame logarithmic spectrum G ; 4) Judge the audio mutation coefficient σ _t and detect the audio waveform mutation and location of mutation regions.

In step 1), the audio signal is transformed to obtain the audio spectrum Y in the logarithmic domain, and the obtained audio spectrum is logarithmically transformed to obtain the logarithmic spectrum G. Specifically, for a digital audio signal y of length h Carry out framing to obtain a matrix whose number of frames is N1 and _whose frame length is 2* N ; add a window function and perform short-time Fourier transform to obtain an audio spectrum Y whose size is N*N1 _; to the audio spectrum Y Perform logarithmic transformation to obtain the logarithmic spectrum G , whose size is N * N ₁ .

Step 2) The energy binarization calculation of the logarithmic spectrum G , specifically, first calculate the maximum value G _max and the minimum value G _min in the logarithmic spectrum G , and set the energy value of each frame frequency as G _ti (1 ≤ t ≤ N, 1 ≤ i ≤ N _l ), the energy binarization value δ(t,i) is calculated by the following formula (1) ,

where λ is the threshold factor.

Step 3) The calculation of the audio mutation coefficient of each frame of logarithmic spectrum G _t , specifically, the audio mutation coefficient σ _t (1≤t≤N ₁ ) is calculated by formula (2) ;

In step 4), the audio mutation coefficient σ _t is judged, and the audio waveform mutation detection and mutation region location are performed. Specifically, it is assumed that the i-th frame G _i and adjacent frames of the logarithmic spectrum G are obtained in step 3). The three audio mutation coefficients σ _i-1 , σ _i , σ _i+1 , if satisfy:

Then it is determined that there is an audio mutation in the audio, wherein the i-th frame G _i is the detected audio waveform mutation region.

Compared with the prior art, the present invention utilizes the property that the signal waveform mutation in the process of tampering and splicing speech causes a short-term energy sudden increase in the frequency of the time segment, and uses a differential algorithm to distinguish between the audio waveform mutation caused by editing and the audio mutation of a strong signal After calculating the audio mutation coefficient, it is detected that the audio has copy-pasted/deleted tampering traces, and the tampering time is accurately located. The method of the present invention has better detection and positioning functions, and can well identify the waveform mutation of the audio when the human ear cannot recognize it and the spectrogram cannot be observed, and it is an effective audio copy-paste/deletion tampering detection technology.

Description of drawings

Fig. 1 is a schematic diagram of inserting audio clips into audio data;

Fig. 2 is a schematic diagram of deleting audio clips from audio data;

Fig. 3 is a schematic diagram of a waveform mutation at a speech tampering splicing point;

Fig. 4 is a detection algorithm flow chart;

Fig. 5 is the oscillogram of detection speech;

Fig. 6 is the logarithmic spectrum of detection speech;

Fig. 7 is the audio abrupt change coefficient of detected speech.

detailed description

Embodiment 1 of the present invention: the detection method of digital audio waveform mutation,

1) Transform the audio signal y to obtain the audio spectrum Y in the logarithmic domain: For a digital audio signal y of length h, divide the audio signal y into frames, and the length of each frame is 2 * N (set N=128); The overlap rate is l (set l = 0.5), then the number of frames N ₁ is:

Let the framed signal be y _i , i=1...N _l , use formula (1) to perform discrete short-time Fourier transform to obtain Y _ti , where w(Nm) is the window function (set as the Hamming window function ).

Use the formula (2) to calculate the spectrogram, and perform logarithmic transformation to obtain the logarithmic spectrum G _i (i=1...N _l ,) , namely

Composing all logarithmic amplitude values into a matrix is the logarithmic spectrum G of the speech signal, and its size is N * N ₁ ;

2) Calculate the energy binarization of the logarithmic spectrum G _i : first calculate the maximum value G _max and the minimum value G _min in the logarithmic spectrum G _i , for the energy value G _i (k) of each frame frequency, k = 1 ... N , calculate the frequency energy binarization value δ _{(t,i) by formula (3)} :

Where λ is the threshold factor (set λ = 0.65), and the matrix Δ composed of δ is defined as the short-term energy binary spectrum. The value of δ _(i,k) is 0, which means that the energy of the kth frequency component in the i -th frame is low, and 1 high energy

3) Calculate the audio mutation coefficient σ _t of the logarithmic spectrum G _i of each frame: due to the artificial copying and pasting of speech fragments, the audio signal will have a waveform mutation at the editing site. This operation introduces new frequency components, making the logarithmic magnitude spectrum contain All frequency energies of a frame with a sudden audio mutation increase suddenly relative to adjacent frames. Therefore, the number of non-zero values of the short-term energy binary spectrum of this frame should be significantly more than that of adjacent frames; in the statistical step 2), each short-term energy binary spectrum The mean value σ _i of the frame is calculated according to the formula (4) to obtain the audio mutation coefficient σ _i (1≤ i ≤ N ₁ );

4) Judging the audio mutation coefficient σ _t , audio waveform mutation detection and mutation area location: due to voice characteristics, strong voice signals will increase the detected audio mutation coefficient, but this signal has time continuity and cannot be copied manually The audio mutation coefficient caused by pasting will only exist in one frame. Therefore, the difference algorithm is used to distinguish the audio waveform mutation caused by editing and the audio mutation of strong signal. The specific steps are to calculate the logarithmic spectrum of each frame by formula (5) Audio mutation coefficient σ _t of G _i (1≤t≤N1) ;

In step 4), the audio mutation coefficient σ _t is judged, and audio waveform mutation detection and mutation region location are performed. Specifically, assuming that in step 3), the i-th frame G _i of the logarithmic spectrum and the adjacent frames are obtained. The three audio mutation coefficients σ _i-1 , σ _i , σ _i+1 , if satisfied:

Then it is determined that there is an audio mutation in the audio, wherein the i-th frame G _i is the detected audio waveform mutation region.

As shown in Figure 5, Figure 5 is a tampered voice waveform diagram after deleting part of the voice, where the time marked by the dotted line is the audio cut position, and the editing trace cannot be discerned by hearing. . As shown in Figure 6, although there is no obvious short-term energy mutation in the logarithmic spectrum of this segment of audio, after calculating the audio mutation coefficient, as shown in Figure 7, it is detected that this segment of audio has copy-paste/delete tampering traces, and Accurately locate the time of tampering.

It can be seen from the above embodiments that the method of the present invention has better detection and positioning functions. In the case that the human ear cannot recognize and the spectrogram cannot be observed, it can well identify the waveform mutation of the audio, which is an effective audio copy-paste/deletion tampering detection technology.

Claims (1)

1. a kind of mutation of DAB waveform detection method it is characterised in that：Comprise the following steps：1) audio signal is converted Obtain the audio frequency language spectrum Y of log-domain, the audio frequency language spectrum of acquisition is carried out logarithmic transformation, obtain logarithm language spectrum G；2) carry out logarithm language Spectrum G energy binaryzation calculates；3) calculate every frame logarithm language and compose G_tAudio frequency change coefficient σ_t；4) to audio frequency change coefficient σ_tCarry out Judge, carry out audio volume control abrupt climatic change and Sudden change region positioning；

Step 1) described in the audio frequency language spectrum Y that audio signal conversion is obtained log-domain, the audio frequency language of acquisition spectrum is carried out logarithm Conversion, obtains logarithm language spectrum G, specifically, carries out framing for the digital audio and video signals y for h for the length, and obtaining frame number is N_l, frame The matrix of a length of 2*N；Add window function and carry out Short Time Fourier Transform, obtaining size is N*N₁Audio frequency language spectrum Y；To audio frequency Language spectrum Y carries out logarithmic transformation, obtains logarithm language spectrum G, and its size is N*N₁；

Step 2) described in carry out logarithm language spectrum G energy binaryzation calculate, specifically, be first calculated logarithm language compose G in maximum Value G_maxWith minima G_minIf the energy value of every frame rate is G_ti(1≤t≤N,1≤i≤N_l), energy is calculated by equation below (1) Amount binary value δ (t, i),

Wherein λ is threshold factor；

Step 3) described in calculating every frame logarithm language spectrum G_tAudio frequency change coefficient, specifically, by formula (2) calculate audio frequency dash forward Variable coefficient σ_t(1≤t≤N₁)；

${\mathrm{\σ}}_t=\frac{\underset{i=1}{\overset{n}{\Σ}}\mathrm{\δ}(t,i)}{n}---\left(2\right);$

Step 4) described in audio frequency change coefficient σ_tJudged, carried out audio volume control abrupt climatic change and Sudden change region positioning, tool Body is it is assumed that in step 3) in obtained logarithm language spectrum G the i-th frame G_iAnd three audio frequency change coefficient σ of consecutive frame_i-1、σ_i、σ_i+1, If meeting：

σ_i>0.85 and | σ_i-σ_i-1|*|σ_i-σ_i+1|>σ_i/16

Then determine in audio frequency there is audio frequency mutation, the wherein i-th frame G_iIt is the audio volume control Sudden change region detecting.