Speaker Discrimination Using Long-Term Spectrum of Speech

Abstract

In this article, we investigate a specific long-term speech spectrum with respect to its use for speaker recognition. The long-term effect was satisfied by averaging short-term autocorrelation coefficients over the whole utterance. The long-term spectrum was calculated by means of second-order linear prediction using the average autocorrelation coefficients. First, speaker discriminability of 32 individual parameters was evaluated by combining spectral energy and spectral slope in eight different frequency bands covering the range 0−4 kHz (seven narrow nonoverlapping subbands and one band spanning over the full range). Then, four subbands with the most discriminative capability were selected for speaker recognition. These subbands involve the frequencies of 0−1.2 kHz in total. In the main experiments, text-independent speaker recognition based on relative Euclidean distance was performed in each single subband as well as in multiple 2 to 4 subbands applying two types of speech data, complete continuous speech and voiced part of the same speech. The voiced speech seems to be generally more effective for speaker recognition using the long-term speech spectrum. The best recognition rates, i.e. 91.7% on complete speech and 100% on voiced speech, were achieved in optimal paired subbands. The long-term speech spectrum can complement the traditional voice features.

Depersonalization of Speech Using Speaker-Specific Transform Based on Long-Term Spectrum

“…It reflects the anatomy of the speaker's vocal tract and can therefore serve as a speaker-specific voice characteristic. This fact was demonstrated in previous experiments where the second-order long-term spectrum was used to identify speakers [18]. For the purposes of transformation in depersonalization, the long-term spectrum is optimized using the LP-spectra 𝑋 𝑘 ( 𝑓 ) of appropriate order which were estimated in all frames.…”