Implementation of Drop-out Concealment Based on Spectral Modelling

The objective of this project is to implement a model based on the Spectral Modeling Synthesis plus noise method, which can be used to continue a time discrete signal. Assuming that a signal is known up to a certain time instant and drops out afterwards. The last valid samples are analysed by computing the short time Fourier transform (STFT), which involves windowing of the signal segment at first (Hanning), followed by the calculation of a discrete Fourier transform (DFT). Subsequently, the parameters of the signal are detected from the magnitude and phase of its DFT-spectrum which is modelled spectrally, i.e., reconstructed by sinusoids and noise. Hereby, the frequency resolution of the reconstruction is determined by the DFT length but can be improved, e.g., by parabolic interpolation. The process of signal continuation is done by preserving the magnitude spectrum with its partial frequencies and amplitudes and computing continuous phase values. This can be seen as continuing the spectrum of the signal, although resynthesis is done in time domain.