Audio Post-Processing
Once the audio signal has been extracted from the images, we
apply some basic post-processing to create the final .wav file.
Resampling
Since the signal is extracted with an effective sampling rate near 64
kHz, we first resample it to 44.1 kHz. Since this a standard audio
sampling rate, it allows our output files to be burned to CDs for
playback on non-computer audio equipment, and it should allow our
output files to be played on all sound cards. Resampling is done with
Matlab's resample function,
which applies an anti-aliasing filter to the original data during
the resampling process, and compensates for the filter's delay so
the waveform is not shifted in time.
Bandpass Filtering
After resampling, we apply a band-pass filter with cutoff frequencies
of 40 Hz and 8000 Hz to remove some extraneous noise. The actual
bandwidth of a 78 RPM record is reported by Olsson et al. to be 40-8000 Hz, and by Stotzer et al. to be 100 - 12000 Hz
(although the latter point out that "It should be noted that for the
most historical records, direct cut disks, the bandwidth is rarely
greater than 6 kHz.")
Playback Equalization
The final step involves equalization for playback. As the original
record was cut, a "pre-emphasis" filter was applied to the audio data
to
account for frequency-dependent variations in the amplitude encoding of
the recording process. To compensate, this filter must be inverted and
applied to the signal extracted from our system. Different pre-emphasis
filters were used (varying with record label and recording equipment),
so we chose to use the generic filter specifications given by Olsson et al.modified slightly for ease of
implementation. Our modified specifications call for 12 dB of gain up
tp
63 Hz, a 6 dB per octave roll-off between 63 Hz and 250 Hz, no gain
between 250 Hz and 3100 Hz, and a 6 dB per octave roll-off above 3100
Hz. The frequency response of the filter is shown below.
Future Work: Noise Reduction
After listening to our audio files, it is quite clear that we need to
focus our future attention on further noise reduction (the average
signal-to-noise ratio on the 78 RPM record extraction is only 11 dB).
Some of this will come from simple changes in the process (e.g. better
cleaning of the scanner and records), some from a more robust
audio-extraction algorithm, and some from post processing. In regard to
the last of these, it may be possible to exploit the noise extracted
from the beginning (or end) of a record where no audio information is
present .
.m files: process_song.m