Improvements in the technology cause not onlydecrease in the sensor costs but also, reduce thedimensions of the sensors. Based on these facts theimportance of signal processing methods increasesday by day. But, in most applications sensors pick upa collection or mixture of signals from many sourcesinstead of only one sensor. In this case, it is hard tounderstand or manipulate the system. For example,recorded EEG signals are mixtures of actionpotentials of many neurons, therefore, the reasons ormechanisms behind epilepsy are still since unknown.For such situations, blind signal separation methodshave been employed. Generally, used blind signalseparation methods are principal componentanalysis (PCA) and independent component analysis(ICA) methods. Although many successful results ofPCA and ICA have been reported in the literature,they suffer from a major drawback; the number ofsources that will be separated must be equal or lessthan the number of sensors. Therefore, it is hard toobtain the true sources of complex signals by thesemethods. As an alternative, the method of delay wasproposed to separate a signal into two independentsources, but its performance is highly dependent onthe delay value.Addition to these methods, artificial neural network(ANN) was also employed for the purpose of blindsignal separation. By ANN, better blind signalseparation results with compared to PCA and ICAwere reported in the literature. In this study, extremelearning machine, which is a novel learning schemeof single hidden layer feed-forward artificial neuralnetwork, was employed. In ELM, the weights andbiases in the hidden layer are assigned arbitraryand the weights in the output layer are calculatedanalytically. Therefore, ELM showed faster trainingstage and higher generalization capability withcompare to back-propagation trained ANN.The proposed approach has three maincontributions, which are: (1) proposed approachcan be employed to separate a signal into a desirednumber of sources without the limitation of PCA andICA, which is the number of sources that the signalscan be separated, must be equal to or less than theobservations, (2) unlike the classical ANNapproaches, in the proposed approach, anorthogonal transfer function was used in the outputof each neuron in the hidden layer, and (3) theproposed approach is extremely fast because of thetraining scheme of ELM.To evaluate and validate the proposed approach,biological signals, which are the EEG, ECG, EMG,accelerometer, gyroscope, and magnetometersignals, were utilized. Achieved RMSE in thetraining stage were in the range of 10-4-64x10-4 andthe duration of training for different lengths of datawere in the range of 0.37-2.33 sec, for separating asignal into 2 and 10 sources, respectively, while thelength of the dataset is 106. Additionally, obtainedcovariance of the separated signals were in therange of 0.2x10-4-38.3x10-4.Obtained RMSE values and fast training stageshowed that the proposed approach has highaccuracy in the separation of a signal and also ithas an extremely fast learning stage. Theseadvantages are because of the training scheme ofELM. Moreover, obtained covariance values showedthat the separated signals are highly independent. Inorder to investigate the reason of obtained lowcovariance, a traditional transfer function wasemployed and not only higher covariance but alsohigher RMSEs were obtained. These results showedthat using orthogonal transfer functions (such assine and cosine, Bessel functions, etc.) after eachneuron in the hidden layer increases the dependencyof the separated signals. Furthermore, two normalEEG and two epileptic EEG signals were separatedinto 1-16 independent sources. Achieved resultsshowed that there is not any relation between thenumbers of separated sources with obtained theindependency of sources of sources and alsotraining accuracy.As a summary, in this study a novel approach, whichis based on ELM, was proposed in order to separatea signal into a desired number of independentsources. Achieved results showed that the proposedapproach is fast and can successfully separate asignal into independent sources.