Noise Suppression Using Gated Recurrent Units and Nearest Neighbor Filtering

A technique to enhance noisy speech through machine learning and digital signal processing methods is proposed in this paper. In the first step of enhancement, Mel-frequency cepstral coefficients are extracted from the noisy speech and fed to a gated recurrent unit (GRU) network which estimates a sequential gain vector used to improve the signal-to-noise ratio (SNR) of the noisy speech. In the second step of enhancement, nearest neighbor filtering is applied to generate an estimate of the isolated noise spectrogram in the noisy speech. This estimate is used to compute a soft mask which is multiplied with the frequency spectrum of the enhanced noisy speech from the first step. This two-step process achieves good results in SNR conditions of greater than 5 db. Under this threshold, the output speech can be distorted. Noisy speech is artificially generated through two datasets consisting of speech, and noise files to create a training and testing dataset for the GRU network.