EMPLOYMENT OF SUBSPACE GAUSSIAN MIXTURE MODELS IN SPEAKER RECOGNITION

EMPLOYMENT OF SUBSPACE GAUSSIAN MIXTURE MODELS IN SPEAKER RECOGNITION

Paper in proceedings (conference paper)

This paper presents Subspace Gaussian Mixture Model (SGMM)approach employed as a probabilistic generative model to estimatespeaker vector representations to be subsequently used in the speakerverification task. SGMMs have already been shown to significantlyoutperform traditional HMM/GMMs in Automatic Speech Recognition(ASR) applications. An extension to the basic SGMM frameworkallows to robustly estimate low-dimensional speaker vectorsand exploit them for speaker adaptation. We propose a speaker verificationframework based on low-dimensional speaker vectors estimatedusing SGMMs, trained in ASR manner using manual transcriptions.To test the robustness of the system, we evaluate theproposed approach with respect to the state-of-the-art i-vector extractoron the NIST SRE 2010 evaluation set and on four differentlength-utterance conditions: 3sec-10sec, 10 sec-30 sec, 30 sec-60 secand full (untruncated) utterances. Experimental results reveal thatwhile i-vector system performs better on truncated 3sec to 10sec and10 sec to 30 sec utterances, noticeable improvements are observedwith SGMMs especially on full length-utterance durations. Eventually,the proposed SGMM approach exhibits complementary propertiesand can thus be efficiently fused with i-vector based speakerverification system.

This paper presents Subspace Gaussian Mixture Model (SGMM)approach employed as a probabilistic generative model to estimatespeaker vector representations to be subsequently used in the speakerverification task. SGMMs have already been shown to significantlyoutperform traditional HMM/GMMs in Automatic Speech Recognition(ASR) applications. An extension to the basic SGMM frameworkallows to robustly estimate low-dimensional speaker vectorsand exploit them for speaker adaptation. We propose a speaker verificationframework based on low-dimensional speaker vectors estimatedusing SGMMs, trained in ASR manner using manual transcriptions.To test the robustness of the system, we evaluate theproposed approach with respect to the state-of-the-art i-vector extractoron the NIST SRE 2010 evaluation set and on four differentlength-utterance conditions: 3sec-10sec, 10 sec-30 sec, 30 sec-60 secand full (untruncated) utterances. Experimental results reveal thatwhile i-vector system performs better on truncated 3sec to 10sec and10 sec to 30 sec utterances, noticeable improvements are observedwith SGMMs especially on full length-utterance durations. Eventually,the proposed SGMM approach exhibits complementary propertiesand can thus be efficiently fused with i-vector based speakerverification system.

speaker recognition, i-vectors, subspace Gaussianmixture models, automatic speech recognition

speaker recognition, i-vectors, subspace Gaussianmixture models, automatic speech recognition

MOTLÍČEK, P.; DEY, S.; MADIKERI, S.; BURGET, L.

2016

19.04.2015

IEEE Signal Processing Society

South Brisbane, Queensland

978-1-4673-6997-8

Proceedings of 2015 IEEE International Conference on Acoustics, Speech and Signal Processing

4445

4449

5

https://ieeexplore.ieee.org/document/7178811

motlicek_icassp2015_0004445