Towards More Efficient DNN-Based Speech Enhancement Using Quantized Correlation Mask

Many studies on deep learning-based speech enhancement (SE) utilizing the computational auditory scene analysis method typically employs ideal binary mask or ratio to reconstruct enhanced signal. However, many SE applications in real scenarios demand a desirable balance between denoising capability and cost. In this study, first, an improvement over attain more superior performance is proposed through introducing efficient adaptive correlation-based factor for adjusting mask. The exploits correlation coefficients among noisy speech, noise clean effectively re-distribute power of during construction phase. Second, make supervised system computationally-efficient, quantization techniques are considered reduce number bits needed represent floating numbers, leading compact model. quantized utilized conjunction with 4-layer neural network (DNN-QCM) comprising dropout regulation, pre-training noise-aware training derive robust high-order mapping enhancement, improve generalization unseen conditions. Results show that outperforms conventional representation other algorithms used comparison. When compared DNN as its learning targets, DNN-QCM provided approximately 6.5% short-time objective intelligibility score 11.0% perceptual evaluation quality score. introduction can weights 5-bit from 32-bit, while suppressing stationary non-stationary noise. Timing analyses also incorporated increase compactness, inference time be reduced by 15.7% 10.5%, respectively.