On a Novel Speech Representation Using Multitapered Modified Group Delay Function

TL;DR

This paper introduces a new multitaper modified group delay function-based speech representation that outperforms existing methods in spectral and cepstral domains, especially in speaker recognition tasks, with comparable computational costs.

Contribution

The paper proposes a novel multitaper modified group delay function (MT-MOGDF) for speech signals, demonstrating superior performance over existing multitaper magnitude estimation techniques.

Findings

MT-MOGDF outperforms MT-MAG in variance and MSE.

Performance is best with Thomson tapers.

Achieves around 20% improvement in speaker recognition accuracy.

Abstract

In this paper, a novel multitaper modified group delay function-based representation for speech signals is proposed. With a set of phoneme-based experiments, it is shown that the proposed method performs better that an existing multitaper magnitude (MT-MAG) estimation technique, in terms of variance and MSE, both in spectral- and cepstral-domains. In particular, the performance of MT-MOGDF is found to be the best with the Thomson tapers. Additionally, the utility of the MT-MOGDF technique is highlighted in a speaker recognition experimental setup, where an improvement of around $20\%$ compared to the next-best technique is obtained. Moreover, the computational requirements of the proposed technique is comparable to that of MT-MAG. The proposed feature can be used in for many speech-related applications; in particular, it is best suited among those that require information of speaker and speech.