DeepMMSE: A Deep Learning Approach to MMSE-Based Noise Power Spectral Density Estimation

Abstract

An accurate noise power spectral density (PSD) tracker is an indispensable component of a single-channel speech enhancement system. Bayesian-motivated minimum mean-square error (MMSE)-based noise PSD estimators have been the most prominent in recent time. However, they lack the ability to track highly non-stationary noise sources due to current methods of a priori signal-to-noise (SNR) estimation. This is caused by the underlying assumption that the noise signal changes at a slower rate than the speech signal. As a result, MMSE-based noise PSD trackers exhibit a large tracking delay and produce noise PSD estimates that require bias compensation. Motivated by this, we propose an MMSE-based noise PSD tracker that employs a temporal convolutional network (TCN) a priori SNR estimator. The proposed noise PSD tracker, called DeepMMSE makes no assumptions about the characteristics of the noise or the speech, exhibits no tracking delay, and produces an accurate estimate that requires no bias correction. Our extensive experimental investigation shows that the proposed DeepMMSE method outperforms state-of-the-art noise PSD trackers and demonstrates the ability to track abrupt changes in the noise level. Furthermore, when employed in a speech enhancement framework, the proposed DeepMMSE method is able to outperform state-of-the-art noise PSD trackers, as well as multiple deep learning approaches to speech enhancement.

On Training Targets for Deep Learning Approaches to Clean Speech Magnitude Spectrum Estimation

“…Deep neural network: ResNet-TCN A modified version of the residual network (ResNet) TCN from Zhang et al (2020) is used to evaluate each training target. 3 The set of hyperparameters for ResNet-TCN used in this work are derived from Zhang et al (2020). It is shown from input to output in Figure 2.…”

“…unit Each block contains three one-dimensional causal dilated convolutional units. Here, we modify the preactivation of the convolutional units in Zhang et al (2020) by using the rectifier linear activation function followed by layer normalisation without the scale and shift operations (again following Xu et al (2019)). The kernel size, output size, and dilation rate for each convolutional unit is denoted in Figure 2 as (kernel size, output size, dilation rate).…”

On Training Targets for Deep Learning Approaches to Clean Speech Magnitude Spectrum Estimation

“…A modified version of the residual network (ResNet) TCN from (Zhang et al, 2020) is used to evaluate each training target. 3 It is shown from input to output in Figure 2.…”

“…The input is first transformed by FC, a fully-connected layer of size d model = 256. Instead of applying layer normalisation (Ba et al, 2016) followed by the rectifier linear function to FC, as in (Zhang et al, 2020), we apply the rectifier linear activation function followed by layer normalisation without the scale and shift operations. This reduces overfitting, as demonstrated in (Xu et al, 2019).…”

“…Each block contains three one-dimensional causal dilated convolutional units. Here, we modify the preactivation of the convolutional units in (Zhang et al, 2020) by using the rectifier linear activation function followed by layer normalisation without the scale and shift operations (again following (Xu et al, 2019)). The kernel size, output size, and dilation rate for each convolutional unit is denoted in Figure 2 as (kernel size, output size, dilation rate).…”