Neil Bryan B. Cazar wrote:
>>3.) it then
passes through a notch filter which filters out the
>>noise. the
noise is the hissing sound you hear when you play cassette tapes
>>or old records.
the output is then converted back to a wav file.
A lot of the noise from old
gramophone records is so-called “surface noise”, cased by small
scratches and detritus jammed into the groove. If you were to look at the
waveform, you would see the smooth music signal interrupted by numerous
spikes. Because these surface noise spikes are short, the noise is very
broadband, and no amount of linear filtering can separate out the signal from
the noise. But non-linear filtering can, in certain cases, do quite a bit
better. I heard a demonstration several years ago of a system that tackled
surface noise by running a linear prediction filter on the signal. The
linear predictor was set to guess what the value of the next sample would be,
and would normally do a pretty good job, leading to a prediction that was close
to the actual sample. The one thing the linear predictor could not
predict was spikes, so when one of those came along, the prediction would
differ a great deal from the true sample. During such periods, the audio
output was taken from the output of the linear prediction filter rather than
from the original programme material. This had the effect of shaving off
the spikes, leaving a smooth waveform behind. I have no doubt that the example
used in the demonstration was carefully chosen to show the technique to its
best advantage, but even allowing for this, the results were remarkable.
I have often thought of trying this sort
of thing out for myself, but have never had the time. Octave would
certainly make an excellent environment for such an experiment.
Steve