Low-variance Multitaper Mel-frequency Cepstral Coefficient Features for Speech and Speaker Recognition Systems

In this paper, we investigate low-variance multitaper spectrum estimation methods to compute the mel-frequency cepstral coefficient (MFCC) features for robust speech and speaker recognition systems. In speech and speaker recognition, MFCC features are usually computed from a single-tapered (e.g., Hamming window) direct spectrum estimate, that is, the squared magnitude of the Fourier transform of the observed signal. Compared with the periodogram, a power spectrum estimate that uses a smooth window function, such as Hamming window, can reduce spectral leakage. Windowing may help to reduce spectral bias, but variance often remains high. A multitaper spectrum estimation method that uses well-selected tapers can gain from the bias-variance trade-off, giving an estimate that has small bias compared with a single-taper spectrum estimate but substantially lower variance. Speech recognition and speaker verification experimental results on the AURORA-2 and AURORA-4 corpora and the NIST 2010 speaker recognition evaluation corpus (telephone as well as microphone speech), respectively, show that the multitaper methods perform better compared with the Hamming-windowed spectrum estimation method. In a speaker verification task, compared with the Hamming window technique, the sinusoidal weighted cepstrum estimator, multi-peak, and Thomson multitaper techniques provide a relative improvement of 20.25, 18.73, and 12.83 %, respectively, in equal error rate.