Abstract:
Frequency modulation has recently emerged as a promising model for
characterising the phase of a speech signal. Proposed is a novel technique
for extracting the frequency modulation (FM) components
from the subband speech signal, using a second-order all-pole
model. Evaluation of a speaker recognition system employing FM features,
extracted using the proposed technique, on the NIST 2001 database
reveals improvement over MFCC baseline and significant
improvements over the discrete energy separation algorithm and a
Hilbert transform based approach in terms of equal error rate.
