Frequency Shift Detection of Speech with GMMs AND SVMs

Abstract

In certain situations, speech might be shifted in the frequency domain amid the presence of noise. To be able to compensate for the spectral shift, it is important to know the amount of frequency shift present. A method based on Mel-frequency-cepstral-coefficient (MFCC) and Gaussian Mixture model (GMM) super vector is proposed for detecting frequency shifts in speech. MFCC or LFCC is extracted to characterize the energy variation of the signal. A GMM is trained for each shifted utterance, and the corresponding GMM super vector is used as the input feature for SVM. Results show that the proposed solution could yield good performance.