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

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