Dimension reduction of the modulation spectrogram for speaker verification.
Tomi Kinnunen, Kong Aik Lee, Haizhou Li · 2008
A so-called modulation spectrogram is obtained from the conventional speech spectrogram by short-term spectral analysis along the temporal trajectories of the frequency bins. In its original definition, the modulation spectrogram is a highdimensional representation and it is not clear how to extract features from it. In this paper, we define a low-dimensional feature which captures the shape of the modulation spectra. The recognition accuracy of the modulation spectrogram based classifier is improved from our previous result of EER=25.1 % to EER=17.4 % on the NIST 2001 speaker recognition task. Index Terms: modulation spectrum, spectro-temporal features, speaker recognition