Neural-based lossy compression of noisy audio signals and their DCT-based post-filtering

DOI: 10.31673/2412-9070.2026.023601

Abstract

A recently proposed method of lossy compression of one-dimensional audio signals is considered with application to musical and speech signals corrupted by additive white Gaussian noise. It is shown that excellent compression ratios exceeding 100 are obtained. Properties of decompressed signals and introduced distortions depend on input signal-to-noise ratio (SNR). For medium ratios, lossy compression is able to partly suppress noise, whilst, for low input signal-to-noise ratios, noise is left. Then, it can be suppressed after decompression by a filter based on discrete cosine transform able to provide improvement by up to 10 dB. Spectral analysis of distortions is carried out and it is demonstrated that the largest distortions are observed for low frequencies.

Keywords: audio signal, noise, lossy compression, DCT-based filtering, distortions.