End-to-end speech enhancement based on discrete cosine transform (1910.07840v4)

Abstract: Previous speech enhancement methods focus on estimating the short-time spectrum of speech signals due to its short-term stability. However, these methods often only estimate the clean magnitude spectrum and reuse the noisy phase when resynthesize speech signals, which is unlikely a valid short-time Fourier transform (STFT). Recently, DNN based speech enhancement methods mainly joint estimation of the magnitude and phase spectrum. These methods usually give better performance than magnitude spectrum estimation but need much larger computation and memory overhead. In this paper, we propose using the Discrete Cosine Transform (DCT) to reconstruct a valid short-time spectrum. Under the U-net structure, we enhance the real spectrogram and finally achieve perfect performance.