On training targets for deep learning approaches to clean speech magnitude spectrum estimation

Estimation of the clean speech short-time magnitude spectrum (MS) is key for enhancement and separation. Moreover, an automatic recognition (ASR) system that employs a front-end relies on MS estimation to remain robust. Training targets deep learning approaches fall into three categories: computational auditory scene analysis (CASA), MS, minimum mean square error (MMSE) estimator training targets. The choice target can have significant impact enhancement/separation robust ASR performance. Motivated by this, produces enhanced/separated at highest quality intelligibility which best found. Three different neural network (DNN) types two datasets, include real-world nonstationary coloured noise sources multiple signal-to-noise ratio (SNR) levels, were used evaluation. Ten objective measures employed, including word rate Deep Speech system. It found estimate priori SNR MMSE estimators produce scores. it established gain ideal amplitude mask scores are most suitable front-end.