CN102074238A - Linear interference cancellation-based speech secrete communication method - Google Patents

Abstract

The invention relates to linear interference cancellation-based speech secrete communication method, which solves a host interference problem in common speech secrete communication technology and improves the decoding performance of secrete information. The method is characterized in that: at a transmitting end, an embedder embeds secrete information in speech predictive residue in a spread spectrum embedding mode; and before embedding, a watermark signal is subjected to pre-cancellation treatment. The watermark-carrying speed signal is transmitted in a common speech channel. At a receiving end, a detector performs linear prediction analysis on the received speech and then performs whitening filtration treatment on the speech signal, and a related detector is used to extract the secrete information. The method can be used in the fields of secrete communication, speed copyright protection, content verification and the like.

Description

A kind of voice conceal communication method of offsetting based on linear disturbance

Technical field

The present invention relates to the multi-media information security technical field, also relate to Digital Speech Communication, Digital Signal Processing in the communication technical field.Specifically, this invention is a kind of Information Hiding Techniques and the digital speech water mark method in the information security field.

Background technology

In the information age, it is more and more frequent that information interchange becomes, and people also need can store safely and transmission information when obtaining information and exchange of information easily.Because the reproducibility of numerical information, it is more and more easier to make to the copyright infringement, distorts also conveniently, has occurred a series of copyright controversial issue problem thus.Information Hiding Techniques is by certain algorithm secret information to be embedded into by in the digital carrier signal (as image, voice, video etc.), proves copyright ownership or realizes secret communication, has become a research focus in the field of multimedia signal processing.It is to be the concealing technology of carrier with the voice signal that voice messaging is hidden, and it has a wide range of applications in fields such as voice copyright protection, secret communication, voice content checking and the recoveries of compress speech bandwidth.

Voice secret communication technology is secret information to be hidden in the disclosed voice signal transmit, and reaches by the transmission that discloses voice messaging and covers the purpose that transmits secret information.Because have the voice messaging of enormous amount in the voice communication channel, the listener-in has just lost concrete target of attack, makes that secret information is difficult to be detected by the listener-in, secret information just is able to safe transmission.For carriers such as image, video, the research that voice messaging is hidden is than later.The voice messaging hidden algorithm has parameter model method, time shaft companding method and spread spectrum hidden method etc. substantially at present.But because voice signal self, these algorithms come with some shortcomings when using.The parameter model method embeds secret information by revising the correlation parameter of model in voice signal.The anti-filtering operation performance of this method is not strong, and it can only use bright detection, has so just limited its range of application.Time shaft companding method is that the duration of some features is carried out the quantification index modulation.This method is not easy modeling, is difficult to control its decoding performance.And existing voice messaging hidden method, all can not the resolved vector voice signal to host's interference problem of watermark signal.Because the power of voice signal is much larger than the power of watermark signal, host's serious interference has influenced the hiding capacity that voice messaging is hidden.Expansion linear disturbance cancellation scheme of the present invention in the frequency covert property of the auditory system that has taken into full account the people, uses the power spectrum of carrier signal watermark signal to be carried out the power spectrum shaping, embed watermark in the prediction residual model of voice signal.And before embedding, watermark signal has been carried out pre-counteracting and handled, utilized then to contain watermark residual error driving linear prediction model synthetic speech signal, eliminated of the interference of carrier voice effectively watermark detection.Solved host's interference problem that voice messaging is hidden, thereby obtained better decoding performance at receiving end.

Summary of the invention

In voice messaging was hidden, voice bearer had a strong impact on the hiding capacity of concealed information.The objective of the invention is at the host's interference problem in the normal speech Information Hiding Techniques, propose a kind of expansion linear disturbance that is adapted to the voice signal carrier and offset conceal communication method.

The characteristics of voice secret communication technology are that the transmission of concealed message must be depended on certain voice bearer.When receiving end adopted blind Detecting, not only interchannel noise had bigger influence to the watermark decode performance, and voice signal also has very big interference to the detection of watermark signal.In voice messaging was hidden, the carrier voice far surpassed the influence of interchannel noise to the interference of watermark detection, and this voice host disturbs and can not be ignored.In the common spread spectrum voice concealing technology, watermark signal is independent of carrier rather than fully is adapted to carrier signal.Carrier signal has been played the part of the role of interchannel noise in decoder end.Be in the Information Hiding Techniques of carrier with the white noise, (Linear Interference Cancellation, LIC) method solves host's interference problem can to adopt the linear disturbance counteracting.Because common LIC method is a kind of method based on coherent detection, coherent detection is actually the projection of received signal on the watermark signal direction and the comparison of detection threshold, and it is the optimum detector of deterministic signal in the white Gaussian noise.Yet voice signal is a kind of coherent signal is the coloured noise signal, and for the coloured noise signal, correlation detector is not optimum.And watermark embedding and detection method that the present invention proposes have solved the host's interference problem in the Information Hiding Techniques of coloured noise carriers such as voice.

The present invention is based on the voice conceal communication method that linear disturbance is offset, the discrete time voice sequence that carrier signal obtains after adopting analog voice signal through the A/D conversion, perhaps other have the coloured noise burst of autoregressive model structure.

In the language secret communication, the watermark deviser wishes under the prerequisite that does not influence the phonetic hearing quality, embeds watermark signal energy as much as possible.According to the masking characteristics of the mankind's auditory system as can be known: when the frequency of two pure tone components was close, off beat can be fallen by forte is hidden, and the people can only feel forte.The method that employing is carried out shaping to the power spectrum of watermark signal, the power spectrum of adjusting watermark signal makes it to have similar shapes with the power spectrum of voice signal, thereby reaches the order of the more watermark signal energy of embedding.For watermark signal is carried out the power spectrum shaping, at first to analyze the characteristic of carrier voice signal.By the generation model of voice signal as can be known: whole spectrum effects of the radiation in the phonation, sound channel and glottal excitation can be reduced to a time varying filter, and its steady-state system function is

$H\left(z\right)=\frac{1}{1+\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{a}_k{z}^{-k}},$

Filter coefficient a wherein _kBe real number, K is the exponent number of model.Voice signal regarded as produce through full pole system model H (z) by list entries.Adopt the LPC method, can obtain the linear predictor coefficients A={a of voice signal _i, i.e. the parameter of all-pole filter H (z).Original watermark signal can realize the power spectrum shaping to watermark signal by being the all-pole filter of filter parameter with the LPC coefficient.

According to the process of watermark power spectrum shaping, set up voice residual error incorporation model.The residual error model description is: original watermark signal is embedded in the speech residual signal, synthesizes the voice signal that contains watermark through vocal tract filter then.Just earlier voice signal is passed through the sound channel inverse filter, obtains speech residual signal, then original watermark signal is embedded into the residual signals that obtains containing watermark in the residual error, pass through the voice signal that filters H (z) obtains containing watermark at last:

$X\left[n\right]=\hat{U}\left[n\right]*h\left[n\right]$

$=\left(W\right[n]+S[n]*h^{-1}[n\left]\right)*h\left[n\right]$

S[n wherein], W[n],

Represent voice signal respectively, original watermark signal contains the residual signals of watermark.h ^-1The impulse response of [n] expression inverse filter, * represents the signal convolution.Method of the present invention is carried out in voice residual error model, has the effect of watermark power spectrum shaping, thereby reaches the purpose of the more watermark signal energy of embedding.

Basic ideas of the present invention and concrete steps are as follows:

A kind of voice conceal communication method of offsetting based on linear disturbance comprises following processing module and step: transmit leg adopts secret information embedding device that secret information is embedded in the carrier voice signal, and the take over party adopts the secret information detecting device to extract secret information;

Secret information embeds device, comprises following processing module and step:

1) according to the secret information sequence, adopt the spread spectrum mode to generate watermark, frequency expansion sequence is selected binary sequence for use and is normalized to the vector of unit length; The unit direction of original watermark signal can represent that the embedding distortion as requested of watermark signal embedment strength is regulated with normalization frequency expansion sequence vector;

2) voice signal of gathering is carried out linear prediction analysis, and obtain the corresponding linear predictive coefficient; Linear predictor coefficient carries out filtering as the parameter of sound channel inverse filter to voice signal, obtains the voice prediction residual signals;

3) original watermark signal and the predicted residual signal that generate are done the pre-processing of offsetting, the watermark signal after obtaining to offset was eliminated carrier and is disturbed before embed watermark;

4) through pre-watermark signal of offsetting after operating, be embedded in the voice prediction residual error; Use contains watermark residual signals driving linear prediction model and generates moisture indo-hittite tone signal; The value of the predictive coefficient of linear prediction model is used step 2) in the linear predictor coefficient that obtains;

The secret information detecting device adopts whitening filtering and detects the form of cascade mutually, comprises following processing module and step:

(1) voice signal that receives is carried out linear prediction analysis, obtain linear predictor coefficient; Use this coefficient as the inverse filter parameter, carry out the albefaction processing, obtain to receive residual signals receiving voice;

(2) according to the watermark template shared with transmitting terminal, use detecting device to detect decoding to receiving residual signals, decoding obtains secret information.

Above-mentioned voice conceal communication method also comprises following treatment step: before the step 4) watermark signal is embedded into the voice prediction residual error, remove the interference of residual signals for watermark detection in advance in watermark signal; Promptly be in watermark signal, to deduct the projection of speech residual signal on watermark signal, the watermark signal after obtaining offsetting:

$\hat{W}=\mathrm{MP\α}-\left(U^{T}P\right)P$

Wherein M represents secret information, and P represents normalized watermark template, and α represents watermark embed strength, (U ^TP) P represents the projection of voice prediction residual error on the watermark direction.

Beneficial effect of the present invention:

Method of the present invention makes full use of the masking characteristics of human auditory system, and on the basis of voice watermark power spectrum shaping, embed watermark information in the voice residual error can reach the purpose of the more watermark signal energy of embedding.Pre-counteracting before the transmitting terminal embed watermark is handled, and has eliminated the interference of voice signal to watermark detection, compares with common spread spectrum voice hidden algorithm, has improved decoding performance.In addition, owing to eliminated host's interference, thus solved the problem of the high correlation between low pass template and the voice bearer, make the low pass watermark to be used, can resist low-pass filtering effectively and attack.And this method attacks for white noise, the speech processes that re-quantization and compression etc. are common and attack and all have robust performance preferably.

Description of drawings

Fig. 1 is that secret information embeds device and detector arrangement figure among the present invention.

Fig. 2 is voice residual error neutral line Interference Cancellation principle of operation figure.

Fig. 3 is the voice signal sample figure that adopts in the embodiment of the invention.

Fig. 4 is the performance map under white Gaussian noise is attacked.

Fig. 5 is the performance map under low-pass filtering is attacked.

Fig. 6 is the performance map under re-quantization is attacked.

Embodiment

Below in conjunction with drawings and Examples concrete embedding of the present invention and detection mode are further introduced.Method of the present invention can adopt microcomputer, digital signal processor (DSP) or special IC to realize according to different scene requirements.

1) watermark embeds scheme

Fig. 1 is that secret information of the present invention embeds device and detection architecture, and wherein telescopiny is as follows:

A. watermark produces: at given embedding distortion restriction D _ePrerequisite under, can adjust the power of watermark signal according to embedding distortion.Concealed message to be embedded can select for use binary sequence M ∈ 1 ,+1}, frequency expansion sequence (watermark template) P selects binary pseudo-random for use and is normalized to the vector of unit length.The direction of original watermark signal can represent that the watermark signal embedment strength is regulated with α with P.

B. the voice residual error produces: according to the residual error model that proposes previously, need carry out lpc analysis to the carrier voice signal, and with the parameter of LPC coefficient as the sound channel inverse filter.Voice signal obtains speech residual signal by the sound channel inverse filter:

$U\left[n\right]=S\left[n\right]+\underset{l=1}{\overset{K}{\mathrm{\Σ}}}a\left[l\right]S[n-l],n=\mathrm{1,2},...L$

{ a wherein _iThe expression LPC coefficient that linear prediction analysis obtains to voice signal.

C. the watermark of voice residual error is offset in advance: Fig. 2 is the schematic diagram that voice residual error linear disturbance is offset operation.In the geometric model of digital watermarking, carrier works and watermark signal can be considered as the vector of higher dimensional space.From geometric model, coherent detection is actually signal that the test side receives in the projection of watermark vector direction and the comparison of detection threshold.The formation of the projection of received signal on the watermark direction is except watermark signal itself also has the projection of carrier signal in the watermark direction.The projection of carrier signal will cause interference to the detection of watermark signal like this.The pre-counteracting operation of the present invention in the voice residual error is when the embed watermark signal, removes the interference of residual signals for watermark detection, the watermark signal after obtaining offsetting in advance in original watermark signal.Just, in original watermark signal, deduct speech residual signal in advance and after the so pre-counteracting operation of the projection on the watermark signal, can obtain watermark signal at built-in end

$\hat{W}=\mathrm{MP\α}-\left(U^{T}P\right)P,$

(U wherein ^TP) P represents the vector projection of speech residual signal in the watermark direction.

D. watermark embedding and phonetic synthesis: the watermark signal that will offset in advance after the operation is embedded in the voice residual error.Adopt the LPC coefficient that obtains among the b, the residual signals that will contain watermark synthesizes the voice signal that contains watermark by vocal tract filter, finishes the telescopiny of secret information.

2) watermark extracting scheme

Fig. 1 is that secret information of the present invention embeds device and detector arrangement figure, and wherein testing process is as follows:

At receiving end, voice signal and attack noise all are the coloured noise signals, and the watermark decode problem is a signal decoding problem in the coloured noise.Optimum detector in this case has prewhitening filter and the correlation detector form of cascade mutually.At first need to use to received signal inverse filter to carry out albefaction and handle, obtain to receive residual signals V.The B parameter of inverse filter={ b _iCan be according to obtaining receiving the speech LPC analysis.

Voice signal has become the white Gaussian noise signal after the liftering, and correlation detector is the optimum detector of deterministic signal in the white Gaussian noise.Because original watermark signal is a polarity reversing signal, and the probability of two kinds of signal appearance is equal, so detection threshold is 0.Detecting device exists

In time, judged Otherwise judge

Adopt such scheme, compare with adopting common spread spectrum voice messaging concealing technology, the superiority of decoding performance is described as follows:

The power of voice signal is much larger than watermark signal and attack the power (generally more than 20dB) of noise signal, and the LPC coefficient A ≈ B of built-in end and test side is at detection statistic T=V ^TOnly remaining interchannel noise distracter and watermark signal among the P, voice or voice residual error are eliminated the interference that watermark signal detects.

In voice communication channel, assailant's purpose is under the prerequisite that does not influence the voice signal acoustical quality, embeds the attack noise of ceiling capacity, to disturb the detection of watermark signal.In order to cover attacking the perception of noise, the selection of its optimum is to compose shaping with the voice signal that contains watermark to attacking the noise power, and the noise signal after the shaping is added to contains in the voice signal of watermark then.Form residual signals at receiving end after liftering, liftering is actually the decorrelation operation, and voice signal and noise signal have become white noise signal.Thereby obtain the decoding error rate of the inventive method:

$P_e=Q\left(\frac{\mathrm{\α}}{{\mathrm{\σ}}_{\hat{N}}}\right),$

Wherein The variance of the white noise of the use before the shaping of expression channel coloured noise.In common spread spectrum voice hidden method, carrier signal is regarded as noise fully, the decoding error rate:

$P_e=Q\left(\sqrt{\frac{{\mathrm{\α}}^2}{{\mathrm{\σ}}_{\hat{N}}^2+{\mathrm{\σ}}_U^2}}\right).$

In general

The performance of this programme will be much better than the performance of general spread spectrum voice hidden method.

According to thinking of the present invention and technical descriptioon, one has voice signal and handles knowledge, the technician of communicating knowledge and computing machine and electronic skill can dispose the specific implementation of each several part according to concrete application scenarios, and can reach the corresponding impact of performance.

Instantiation: gather one section voice signal, implement secret information according to such scheme and embed and leaching process.The carrier voice signal that Fig. 3 is to use, sampling rate are 10kHz, and frame length is 20ms.The frequency of frequency expansion sequence is 2kHz.Per 200 sampled values are embedded in the 1bit watermark information in voice signal.For detecting the robustness of secret information in the such scheme, the voice signal of embed watermark is done following attack handle, and relatively this scheme and the generally performance of spread spectrum voice messaging concealing technology:

1) superimposed noise: with the signal to noise ratio (S/N ratio) of-20dB～0dB scope, white Gaussian noise superposes on the voice signal of embed watermark.Fig. 4 is this programme and the performance of general spread spectrum voice messaging hidden method under white noise is attacked.

2) compress speech: the voice signal that will contain watermark respectively carries out the compression of following different modes: IMA ADPCM, CCITT A-law and CCITT μ-law compression.Voice after will compressing then carry out 10kHz, 16, monaural PCM compression again.Table 1 is for embedding the decoding performance result of 100 bit watermark information under following different compress modes.

Watermark decode performance under the table 1 language compression attack

3) low-pass filtering: adopt the low-pass filter of cutoff frequency 1kHz～4.5kHz, to the voice signal low-pass filtering of embed watermark.Watermark signal results from the low pass watermark template of white noise watermark template and 2kHz respectively.Fig. 5 is that low pass watermark and plain boiled water are imprinted on the performance of low-pass filtering under attacking in this programme.

4) re-quantization: the voice signal of embed watermark is carried out 0bit～16bit quantification again.Fig. 6 is this programme and the performance of general spread spectrum voice messaging hidden method under quantification is attacked.

As can be seen, method of the present invention is compared with common spread spectrum voice hidden method, has improved the decoding performance of secret information, and this method all has stronger robust performance for above several frequently seen speech processes and attack.

The method of the invention can be used for fields such as secret communication, voice copyright protection, content verification.

Claims (2)

1. voice conceal communication method of offsetting based on linear disturbance, it is characterized in that comprising following processing module and step: transmit leg adopts secret information embedding device that secret information is embedded in the carrier voice signal, and the take over party adopts the secret information detecting device to extract secret information;

Secret information embeds device, comprises following processing module and step:

1) according to the secret information sequence, adopt the spread spectrum mode to generate watermark, frequency expansion sequence is selected binary sequence for use and is normalized to the vector of unit length; The unit direction of original watermark signal can represent that the embedding distortion as requested of watermark signal embedment strength is regulated with normalization frequency expansion sequence vector;

2) voice signal of gathering is carried out linear prediction analysis, and obtain the corresponding linear predictive coefficient; Linear predictor coefficient carries out filtering as the parameter of sound channel inverse filter to voice signal, obtains the voice prediction residual signals;

3) original watermark signal and the predicted residual signal that generate are done the pre-processing of offsetting, the watermark signal after obtaining to offset was eliminated carrier and is disturbed before embed watermark;

4) through pre-watermark signal of offsetting after operating, be embedded in the voice prediction residual error; Use contains watermark residual signals driving linear prediction model and generates moisture indo-hittite tone signal; The value of the predictive coefficient of linear prediction model is used step 2) in the linear predictor coefficient that obtains;

The secret information detecting device adopts whitening filtering and detects the form of cascade mutually, comprises following processing module and step:

(1) voice signal that receives is carried out linear prediction analysis, obtain linear predictor coefficient; Use this coefficient as the inverse filter parameter, carry out the albefaction processing, obtain to receive residual signals receiving voice;

(2) according to the watermark template shared with transmitting terminal, use detecting device to detect decoding to receiving residual signals, decoding obtains secret information.

2. the described a kind of voice conceal communication method of offsetting of claim 1 based on linear disturbance, it is characterized in that also comprising following treatment step: before the step 4) watermark signal is embedded into the voice prediction residual error, in watermark signal, remove the interference of residual signals in advance for watermark detection; Promptly be in watermark signal, to deduct the projection of speech residual signal on watermark signal, the watermark signal after obtaining offsetting:

$\hat{W}=\mathrm{MP\α}-\left(U^{T}P\right)P$

Wherein M represents secret information, and P represents normalized watermark template, and α represents watermark embed strength, (U ^TP) P represents the projection of voice prediction residual error on the watermark direction.

Images