COST278 and ISCA Tutorial and Research Workshop (ITRW) on Robustness Issues in Conversational Interaction

University of East Anglia, Norwich, UK
August 30-31, 2004

Robust SNR Estimation of Noisy Speech based on Gaussian Mixture Modeling on Log-Power Domain

Tran Huy Dat, Kazuya Takeda, Fumitada Itakura

CIAIR, Nagoya University, Nagoya, Japan

This work presents a blind SNR estimation based on Gaussian mixture modeling (GMM) of observed noisy speech on the log- power domain. By describing SNR measures as the expectation of a function of two random variables of local noise and noisy speech powers on the log-domain, their distributions can be estimated via the EM algorithm. Once the parameters of distributions are estimated, SNR measures can be derived statistically. Both of segmental and global SNR estimations are derived in this work. The experimental results show the effectiveness and robustness of proposed method for different types of noise.


Full Paper

Bibliographic reference.  Dat, Tran Huy / Takeda, Kazuya / Itakura, Fumitada (2004): "Robust SNR estimation of noisy speech based on Gaussian mixture modeling on log-power domain", In Robust2004, paper 15.