Global SNR Estimation of Speech Signals using Entropy and Uncertainty Estimates from Dropout Networks

TL;DR

This paper introduces two innovative methods leveraging dropout-based uncertainty and entropy from DNN-HMMs to accurately estimate the global SNR of speech signals across various noise conditions.

Contribution

It proposes novel SNR estimation techniques using dropout-based uncertainty and entropy measures within DNN-HMM speech recognition models, enhancing noise robustness.

Findings

Effective SNR estimation across diverse noise profiles

Utilizes dropout-based Bayesian uncertainty for improved accuracy

Demonstrates robustness at SNR levels from -10 dB to 30 dB

Abstract

This paper demonstrates two novel methods to estimate the global SNR of speech signals. In both methods, Deep Neural Network-Hidden Markov Model (DNN-HMM) acoustic model used in speech recognition systems is leveraged for the additional task of SNR estimation. In the first method, the entropy of the DNN-HMM output is computed. Recent work on bayesian deep learning has shown that a DNN-HMM trained with dropout can be used to estimate model uncertainty by approximating it as a deep Gaussian process. In the second method, this approximation is used to obtain model uncertainty estimates. Noise specific regressors are used to predict the SNR from the entropy and model uncertainty. The DNN-HMM is trained on GRID corpus and tested on different noise profiles from the DEMAND noise database at SNR levels ranging from -10 dB to 30 dB.