Speech Enhancement Algorithm Based on a Convolutional Neural Network Reconstruction of the Temporal Envelope of Speech in Noisy Environments

Temporal modulation processing is a promising technique for improving the intelligibility and quality of speech in noise. We propose enhancement algorithm that constructs temporal envelope (TEV) time-frequency domain by means an embedded convolutional neural network (CNN). To accomplish this, input signals are divided into sixteen parallel frequency bands (subbands) with bandwidths approximating 1.5 times auditory filters. The corrupted TEVs each subband extracted then fed to 1-dimensional CNN (1-D CNN) model restore distorted method evaluated using 2,700 words from nine different talkers, which mixed speech-spectrum shaped random noise (SSN), babble noise, at signal-to-noise ratios. Short-time Objective Intelligibility (STOI) Perceptual Evaluation Speech Quality (PESQ) metrics used evaluate performance 1-D algorithm. Results suggest improves STOI scores on average 27% 34% SSN respectively, PESQ 19% 18%, compared unprocessed speech. also shown outperform conventional TEV-based