Balakrishna Lavu; Ashokkumar P V MEE10:114, pp. 42. ING/School of Engineering, 2011.
In this thesis, we proposed an approach for extracting a desired speech signal from a mixed source signal using the ICA-R algorithm and Bessel features. Here the desired
speech signal and the reference signal are two different speech utterances of a same speaker.
In the current existing literature most of the methods deal with designing the reference signal with prior information of the target speech signal. The crucial problem is the design of reference signal in advance which is close to the desired signal when the desired source signal is very week in mixed signals and also when there is no prior information about the desired source signal.
In the proposed method we do not require any prior information about the desired speech signal that has to be extracted. The ICA-R algorithm is extended to use Bessel coefficients of the observed signals and the reference signal for processing as they are more efficient in representing speech-like waveform.
From the simulation results and the performance analysis in chapter 4, comparing the proposed method with one of the previous existing methods shows that the proposed
method is more effective. This shows that the computation done at the feature level i.e. the Bessel coefficients of the signals yields better results than on the sample values.
This is very useful for many applications such as speaker verification, speaker identification and so on even when the desired signal is very week in mixed signals. An example of real world application where the proposed method is useful is the voice recognition and speaker verification login even in the presence of external disturbances and also in voice security applications.