Signal Reconstruction from Multiresolution Magnitude Spectrograms for Audio Signal Processing

The reconstruction of missing phase information is an important issue for the reconstruction of audio signals with exclusive knowledge of their magnitude spectrograms. However, the quality of the reconstructed phases often suffers from a trade-off between time and frequency resolution. Depending on the preset window length, this trade-off leads to widely known artifacts like transient smearing and bass distortions, which can be greatly reduced by using multiple window lengths.

In this thesis, a new approach for signal reconstruction from magnitude spectrograms with multiple window lengths is developed and evaluated. The new technique is extended to stereo signals, where the phase differences between the left and right audio channel are important to keep the stereo image intact.

The approaches are evaluated on time-stretching and pitch-shifting tasks. As a new application, this thesis shows how the combination of magnitude spectrogram processing and phase reconstruction can avoid the comb-filter effect in audio mixing.