Cancellation Of Background Noise In Speech Communication Computer Science Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

In many speech communication applications, the presence of background noise causes the quality or intelligibility of speech to degrade. This project is trying to study a method to cancelation of background noise in speech this project involves investigating the effectiveness of independent component analysis for the purpose of noise cancellation and also the relative usefulness of this approach . The project needs good understanding of digital signal processing and familiarity with MATLAB. In the Noise cancellation module Independent Component analysis is adopted to remove the noise from the original produced signal to give a better quality and intelligibility to it.

The major defect in the speech processing is inteference of noise in between the speech signals. The signal strength and accuracy is reduced due to these interference of these noise. These noices are produced due to the traffic, crowds, ventilation utensils and electrical signals as well.

Literature review:

2.7 Speech production:

Basically most of the speech signals can be classed as voiced or unvoice. In general terms,the source that can create a quasi-periodic pulse waveform or a random noise waveform. In the case of voiced speech, the impulse train generator produces a sequence of unit impulses, these impulses are located by the desired fundamental period. Ths signals in turn excites a linear systems,these impulse response have the desired shape. The gain control mechanism controls the intensity of excitation .

The sources of excitation generates impulse,In the out side world which is controlled from, the pitch-period signal and a random number generator. The pitch period is reciprocal to the thes signal frequency .It represents the velocity of oscillations of the vocal cord. In the case of unvoiced signals, TheRandom number generator output simulates the pressure buildup waveform and the quasi-random turbulence .

The vocal tract is in the fixed formation, it is modelled as a 'linear time invariant system. The speech is convolution of the impulse response of the vocal tract by the excitation waveform. But in practically it is not possible, because the vocal tract shape changes to generate different sounds and we know that the vocal tract shape changes from person to person. For that reason in practical the vocal tract is designed as a 'linear time varying' digital filter. But in the process of speech production, few minutes the vocal tract configuration is fixed based on this nature of the vocal tract, we divide the speech waveform into frames of duration of few milliseconds each, after this do the processing.

The digital filter simulates the vocal tract system, in that process the filter coefficients specify in some method. During the continuous speech process the vocal tract as a function of time.

The gain control over between the source and system. The system allows certain flexibility levels of acoustic input. The filter output corresponds to the final level of speech output at the appropriate rate.

The main aim of speech analysis systems is, to appropriate model parameters are estimated

From the real speech at any reasonable conditions, to derive synthetic speech signals that one is same as the original signal.

Signal processing represents the signal based on a given model and the application of some upper level transformation in order to put the signal into a more expedient form. The extraction and the utilization of the information is the final steps in this process. Hence the process of speech production consists of two tasks,

First one is, it is a vehicle to contain a general illustration of a speech signal in parametric form or waveform.

Second, signal processing serves the function of aiding in the process of transmission the signal demonstration into alternating forms which are less general in nature, but more appropriate to specific applications.

well known sampling theorem guided the reprsentation of speech signals ,this theorem states that a band limited signals is represented by samples taken periodically in time, the samples are taken at a very high rate.

The representations of parametric on the other hand is concerned with representation of the speech signals. The parameters of this model are conveniently classified as excitation parameters or vocal tract response parameters .Speech processing is also related to natural language processing (NLP), in this procesas its input can come from or output can go to NLP applications.

Speech processing can be divided into the many categories like speech recognitation , speaker recognition,speech enhancement excetra.

2.8 Speech Enhancement:

In the past days speech enhancement has used for suppression of additive background noise. additive noise is easy to deal with convolutive noise or nonlinear disturbances.[26]

The main aim of speech enhancement is to improves the quality and intelligibility of degraded speech signal by using different algorithams and audio signal processing techniques.Noise or noise reduction factors degrades the speech signals.Speech enhancement Is the most important tehnic speech processingIt is used for many applications like, mobilephones, teleconferencing systems and speech recognition.

2.9 Noise cancelation:

Noise is a unwanted sound,it is loud and irritating.Noise is interfere in hear speech signals and other required sounds. Noise can slow down the ability to concentrate and cause the health problems for people who are work in the noisy environments. Even students also like to relief from noisy environments when they are trying to concentrate on their studies. Noisy environments is also make it difficult to listen to music with a earphonest.

Noise creates many problem to our health and safety, interferes with communication, and reduces our ability to enjoy life at its fullest. To the noise control through passive measures has been a systematic effort. while high frequency noise and vibration is also agreeable to passive controls, it has confirmed that the difficult to passively control low frequency noise and vibration. IN the past days the active noise cancelation is proposed for low frequency noise signals but it was impratical at that days. These days it has a practical solution to many previously difficult problems in environmental noise.

2.10 How Noise cancelation works:

Niose is a sound signal, which is loud and harmful.It generate from roaring of engines clanking of machinery.We no need to listen the noise.

All signals having waveforms.simple illustrations are the waveform of vibrating the guitar or the waves in a pool of water.[27]

2.11 Adding waves:

If we add two wave forms, then they are moving in same direction .these signals are in phase ,that means the peaks and values of the signals are scheduled.The amplitude of the waves is doubled and the sorces are same sound.It sound will be doubles the volume.

2.16 cancelling waves:

If we add two wave forms, then they are moving in same direction .these signals are in out of phase.the amplitude of the waveforms cancel each other.

Waves out of phase will cancel each other

From the above figure,the adding of two waves will gives flat wave.In this conditions we would here no sound.

2.17 Doing it with electronics:

Canceling the sound waves can be done by electronically. There are special noise suppression hearaids,it consists of a microphone and electronics built in. The microphone identifys the noise,after that changes that one into an electrical signal and moves it to the speaker in the headset, it turns the signal back into sound. In this way microphone-speaker system works.

This process makes it different, that the electronics place the actual signal exactly out of phase with therecorded signal, this one is too loud , it easily passes through headset to the ear. The sound from the headset then is just the same sound and as loud as the noise, but it is completely out of phase with the noise, then it cancels the sound signals.

2.18 Active noise cancelation:

Active noise cancelation is a noise cancelation method,which is used to decrease or totally cancel the unwanted sound.Then we cannot hear it.In this process electronics worked actively, it causes the noise cancelation I real time.If we want cancel the sound ,in the case of single sound frency.We have to add the signal with the frequency of 180 degrees out of phase.It is more difficult with complex soundsignals.the most of the electronic noise cancelation devices contains special earphones.other type of devices not needed earphones. [28]

Practical application of this process l has to stay as the electronic technology available at that time was not sufficient for execution of Active Noise Cancellation systems. These days digital technology is used to the point where the cost effectual DSP microcomputers perform the intricate calculations exist in the noise cancellation process. Advance features of this technology has slove the problems which is not sloved in the low frequency environmental noise at a reasonable cost.

The fast development in DSP technology for the processing of discrete, real-time signals has broadened the range of various noise suppression systems for the speech enhancement. Others type of adaptive systems based on signal processors introduced new and talented area of applications due to low installation and operation costs. Then, the adaptive system , which is based on the DSPprocessors have improved the performance with existing applications .

. in addition, DSP technology development has give a space for new methodology implementation and principles, already designed but due to extremely high installation costs and computational complexity and the insufficient processing power.

2.19 Fast fixed point type algorithm:

A fast fixed-point(Fast FPA) type algorithm is used for separating the complex values and linearly mixed signals ,which is obtainable and its computational efficiency is shown in simulation. The estimator gives the local consistency of the algorithm is proved. In the FPA computations are done in blockmode.In single step of the algorithm large number of data points is used.FPA is also used for blind source signal separation and feature extraction.

Fixed-point algorithm:

In the case of FPA, for complex signals under the condition of ICA data model, which searches for the extrema of E{G(|wHx|2)}.

This algorithm needs a first whitening of the data the observed values are mentioned below:

xold= linearly transformed to a zero-mean variable

Here x= Qxold

x= (x1r+ix1i……, xnr+ixni)

Therefore E {xxH} =I.

The FPA for one unit is

W+=E{x (wHx)*g (|wHx|2) +│wHx│2g'(|wHx|2)}w

wnew = w+ / ||w+||

The one unit algorithm is wide ranging for estimation of the total ICA transformation (s= WHx).to stop different neurons from the converge to the same maxima. The output is w1Hx…wHnx, there are decorrelated after the every iteration. This deflation scheme based on Gram Schmidt decorrelation scheme. If we have estimated p independent components, or p vectors w1…wp, we execute the one-unit FPA for wp+1 the projections of the previously estimated p vectors and then renormalize wp+1.[6]

The above mentioned decorrelation process is suitable for deflationary division of the independent components. Sometimes it is used for estimate all the independent components simultaneously and use a symmetric decorrelation. This can be mentioned below. [6]

W= W (WHW)-1/2

Where W = (w1….wn) is the matrix of the vectors.

2.20 Blind signal separation algorithm:

Blind signal separation(BSS) is a algoritham,which is used to separate the set of signal from a set of mixed signals, without taking the support of information or the mixing process.BSS is based on the assumption of that the source signals, its do not correlate each other. Suppose, if we consider the signals, these signals may be mutually statistically not dependent or decor related. BSS separates the set of signals into a set of other signals, therefore regularity of each signal resulting signals is maximized, and the regularity is minimized between the signals , that means these signals statically independence is increased, because the resulting signals temporal redundancy is clamped .The resulting signal are more effectively deconvoluted than original signals.[17]

ICA is important for BSS and important for more practical applications. This one is related to the search of a factorial of the data, that means a new data vector value represents the each of the data vector therefore it gets uniquely encoded by the resulting vector but the data vector code components are statistically independent

The above figure shows the planned selective noise control system with two microphones using BSS algorithm. Let us assumed to be that two independent source signals are voice and noise signals.

The two microphones receives the mixture of two independent source signals through the space.

We have to keep the voice signal around microphone, while it is selectively eliminating the effect of noise.

The selective noise control system [17] Fig. The dynamic recurrent neural model

Net work with two Microphones [17]

Fig shows the proposed selective noise control system with two microphones using BSS. Let it be assumed that two independent source signals are voice and noise. The mixtures of two independent source signals are received at two microphones through the space. We have to retain only the voice signal around microphone 2while selectively eliminating the effect of noise [17].

Usually the BSS is unknown for the order of the separated signal, but it is possible to recognize the variation order of the BSS output signals using a simple heuristic or statistical method such as a evaluation of the kurtosis of the separated signals. Separated noise signals are passes through the adaptive filter of the ANC system.

2.24 The problem of whitening

Whitening is the useful pre-processing technic in ICA,it is to first whiten the observed variables. It means before the application of the ICA algorithm.

If we take the random vector x with n number of elements, we have to take a linear transformation V into another vector z therefore it is a white or shaped.


Here E=[e 1.......... e n ] , is a matrix,its columns are the unit.

C x =E{xx T} it is a eigenvectors of covariance matrix

D=diag[d 1 .........d n ] is the diagonal matrix of the eigenvalues of C x

Therefore C x =ED E T .This process is called the eigenvectors decomposition of the covariance matrix.

Below The linear whitening transform is expressed as

V = D-1/2 ET

Hence V = D-1/2 ETx

ICA estimation is applied on the whitened data z, in place of the original data x. the whitened data is adequate to find an orthogonal separation of matrix, if the independent components are assumed to be white. Dimensionality reduction by PCA is carried by projecting the N number of dimensional data to the lower dimensional space spanned by m (m x ). The eigenvectors matrix E dimensions is N-m and the diagonal matrix of eigenvectors D dimension is m - m . basically, this one is a nontrivial field ,which is to classify the lower dimensional subspace properly. For un noise data, a subspace corresponding to the non­zero eigenvalues is need to be found.

Fig(a) : Illustration of mixing and separation system. (A) is the mixing matrix and (B) is the unmixing matrix[19]

In many conditions ,the data are contaminated by noise and this is not contained exactly within premises of the subspace. In this conditions if we describe the data , the eigenvectors resultant to the largest eigenvalues. Generally,the weak independent components may be lost in the dimension of the reduction process. This is consists of hit and trial method. Dimensionality reduction can be determined by the methods other than PCA. This methods consists of local PCA and random projection mechanisam. fig (a) represents a two step process,that is whitening and




Figure (b) is a Schematic of separation[20].

the data is whitened automatically the matrix W is orthogonal. This one decreases the number of parameters to be estimated and enables the use of efficient optimization techniques.

2.24Eigen values and Eienvectors

According to the mathematics, eigenvalues and eigenvectors are related topics of linear algebra.these are the properties of a matrix. These computed method are mentioned below. Eigenvalues and eigenvectors gives more useful information about the matrix and it is used in Matrix factorization technics. These applications are used in mathematics, economics and quantum mechanics.

Generally, If we apply matrix on vector by changing both its magnitude and direction. Suppose,a matrix may works on perticular vector ,its changing their magnitude and leave theit direction as it is. This type of vectors called eigenvectors of matrix.Multiplying vector magnitude to matrixs,this one act as a is positive,when its direction is unchanged other wise it would be negative. This value is called eigenvalue associated with that eigenvector. An eigenspace is a set of all eigenvectors,its having the same eigenvalue, together with the zero vector.[21]

These concepts are not formally defined without the prerequisites, including an information of matrices, vectors, and linear transformation. The technical contents are mentioned below.

Generally, if A is a linear transformation, a non-null vector x is an eigenvector of A if there is a scalar λ therefore

Where λ= Eigen value

A= matrix

X =eigenvector

2.25 Fast ICA algoritham:

Fast ICA algoritham is a computationally highly efficient technic for performing the estimation of ICA. It is used in a fixed-point iteration process,which has been found in independent process.It is 10 to100 times showing faster performance than the conventional gradient descent technic for ICA. one more additional advantage of the FastICA algorithm is, it is used for perform projection pursuit. therefore it is providing a general purpose data analysis method,it is used for both an exploratory fashion and estimation of independent components.[18]

2.24 Preprocessing of Data for ICA

Usually, ICA is used on multidimensional data. It may be ruined by noise, and many original dimensions of data may contain only noise. This is the reason for the poor results of high dimensional data as it contains very few latent components. That is the reason why reduction of the dimensionality of the data is carried before ICA. Therefore, the noise is reduced by knowing a principal subspace where the data exists. Apart from these when the parameters are too many when compared to the number of data points, predicting these parameters becomes very difficult and may result in over-learning. This may leads to prediction of the independent components that contain unit spike or bump and are usually zero anywhere.

. The reason is that the space of source signals of unit variance, nongaussianity is more or less maximized by such spike/bump signals. Inastead of minimising the dimension, the obtained signals arecenteredanddecorrelated.The observed signal X is centered by subtracting its mean:

Second-order dependences are deleted by decorrelation, which is obtained by the principal component analysis (PCA) . The ICA problem is simplified if the observed mixture vectors are first whitened. A zero-mean random vector z =(z i .....z j ) T is said to be white if its elements z are uncorrelatedandhaveunitvariancesE{zizj}=δi,j.

In terms of Covariance matrix, the above equation is given as,

where I is the identity matrix. A synonymous term for white is spread. If the density of the vector z is radially symmetric and appropriately scaled, then it is sphered, but the converse is not correct all the time, as whitening is necessarily decorrelation followed by scaling, for which the PCA technique can be adopted[12].


3.1 Architectural Design

The below figure shows the relationship between the different components of system. This figures gives the understanding of the total concept of system

Separated Signal

Source Signal

Noise l

Noise 3

Noise 2

Block Explanation:

The first block consist of source signal, in the second stage the signal is mixed with noise signal like randomized signal the following stage the Eigen vectors will be drawn for each signal with the help of pca method. After this stage the noise is removed from the mixed signal. Finally, the original sound is extracted as the output signal by using ICA.


For technical computing mat lab is the best programming language, because it is a high performance language. Mat lab software improves the calculating efficiency and use the program in easy access to matrix software. This one is developed by linear system project and Eigen system projects.Matlab is designed on the basics of highly developed matrix. In this basic element is matrix. This will not require the predimension.

The use of mat lab , Mat lab is used for algorithm development ,data analysis ,scientific fields and engineering streams and so on.

The package selected to developed noise cancelation project is mat lab 7.0.This one is developed in image processing demoing.

Mat lab system Mat lab system having five main parts, in that fields first one is development

Field, it consists of set of tools which can help to use mat lab functions and files.

The second field is mat lab mathematical function field, which is having the large numbers of computational algorithms. Its range from basic functions to critical functions.

Third field is mat lab language, which is high level matrix programming language and control flow statements.

Starting matlab, it will be run on the windows platform and unix platform.

3.3.Procedure steps for background noise cancelation by using ICA:In the first we will take the input source signal and generate a gaussian noise ,funky and triangular signals,after that divide the signals by using standard devation.Aater these steps we have to remove the mean in the all signals,finally create mixture of the noise signals.after doing all these steps, compute the PCA.In the PCA process whitening the data vector and dewhitening signal matrix.This is done by using eigen value decompositionAt the end of the process perform ICA,after this process we will get the clear speech signal.

4 Module description :

4.1 Eigen vector Transformation :

The mpodule uses traing data from all speech signals, that means noise to calculate the covariance matrix ,it uses this matrix to find the eigenvalue and eigenvector matrix to perform noise seperation.

4.2. Noise cancelation module:

In this noise cancelation module ICA techinique will be adopted to remove the noise from the produced original signal

5 conclution:

The main development of this project is implemented by ICA with mat lab software .Fast fixed -point algorithms are used for separating the noise signal and blind algorithms are used for separating mixed signals. Fixed-point algorithm improves the performance of the clean speech. ICA is adopted to improve the noise from original produced signal to give a better quality speech signals. The project titled is successfully completed with the implementation of ICA with more accuracy than the present techniques

Screen shots

Main Screen

Original Signal

mixed signal

eigen Values

separate signal

Clear speech signal

6.Future work:

The best technique available in the present trend for cancelation of noise in speech is ICA. If there is any advancements in the present technology more sophisticated methods are also adopted for the cancelation of noise in speech