Extraction of a Speech Signal in the Presence of a Musical Note Signal

Speech and musical note signals have overlapping spectra. Hence, extraction of the speech signal cannot be done by using conventional filters. The new approach is to employ a type of artificial neural network (ANN) called the General Regression Neural Network (GRNN). The GRNN does not need any information from the frequency domain – it learns to map an input signal to the desired output signal by way of examples. The methodology involves training the GRNN to map an incoming signal, which is a composite of speech and musical note signals, to the desired output signal, which is the musical note signal alone. If the mapping is done successfully, then the speech signal can be extracted by subtracting the output signal from the input signal.