Speaker Identification Using Robust Speech Detection and Neural Network 1
Atanas Ouzounov · 2007
An experimental study of the effectiveness of two speech detection parameters in the text-independent speaker identification task is presented in the paper. The first parameter is obtained by processing the spectral autocorrelation function derivative, while the second one is based on the multi-band spectral entropy. The techniques employed are: the two above mentioned parameters and a single MLP for speech detection, LPC cepstrum as a speaker identification feature and a common (for all speakers) MLP for speaker classification procedure. The training and testing have been done using noisy telephone speech data from BG- SrDat corpus. The experiments have shown that in comparison with the multi-band spectral entropy, the use of the spectral autocorrelation function derivative in speech detection results in a lower speaker recognition error.