An optimal frequential based splitting approach in signal restoration: the mixed-filter-bank, realisation and validation

In the area of music signal restoration, methods based on the short-time-Fourier-transform (STFT) are usually used. They yield good results when the local signal to noise ratios (SNR) are higher than 10 dB on the whole band. Most of time, this condition is not respected and a poor detection of musical components is engendered. In the paper, a mixed filter bank approach is proposed The signal is first split in a Δf/f filter bank by means of a wavelet transform and then a noise reduction method based on STFT is used with different window lengths in each subband The choice of parameters in these subbands is connected to the local SNR. Listening tests have shown a real audible improvement of this method applied to an old musical recording