re: the restoration question....

[St-Laurent Gilles <gsl@nlo.nlc-bnc.ca>]

> (1) What exactly do CEDAR and NoNOISE do?

The following is an excerpt of an article I wrote about restoring 7" 
Berliner (the first flat discs made) for our upcoming web site "The Virtual 
Gramophone".  The site with the complete article including audio examples 
should be up in a few weeks.


[...]
4. Digital noise reduction
Three general classes of noise are found on sound recordings: clicks, 
crackle and hiss. CEDAR (Computer Enhanced Digital Audio Restoration) 
removes or reduces these imperfections with the De-Clicker DC1, the 
De-Crackler CR1 and the De-Hisser DH2. These units are based on twin 40-bit 
floating-point processors that process sound in real time (i.e. there is no 
waiting while the units are calculating the results).
4.1 De-Clicking
The benefit of removing high frequency, high-energy transient noises, such 
as clicks and pops, becomes immediately apparent. The DC1 removes both 
clicks and any underlying music. It then re-creates the missing sound wave 
by analyzing pre- and post-click samples and interpolating the results using 
high order algorithms. The number of samples that the DC1 examines depends 
on the length of the click. A short click requires fewer samples (10) than a 
longer pop (60 to 200) to rebuild the sound wave. The De-Clicker can remove 
up to 2500 clicks per second per channel in real time.
4.2 De-Crackling
Crackle is a burst of short, small spikes which is added to the original 
sound by poor record surface quality, buzzing caused by improperly wired or 
grounded equipment, or distortion caused by overloading amplifier mixer 
outputs or digital clipping. These introduce a harshness to the sound. 
Crackle is a more subtle and difficult form of noise to remove than a click. 
The De-Crackler addresses this problem by dividing the input signal into 
"genuine" signal and "crackle/distortion" signal, and working solely on the 
signal with crackle. First, the operator determines the the level of 
"crackle/distortion" present in the input signal, then adjusts the amount of 
crackle that the CR1 is required to remove. Next, the Crackle Mode is set to 
either Crackle 1 (sharp and well defined) or Crackle 2 ('grungy' and not so 
well defined). Finally, the signal is recombined. The CR1 must be adjusted 
by ear and, if not set correctly, can have a detrimental effect on the sound 
quality .
4.3 De-Hissing
Hiss is quite obvious to a human listener but is far more difficult for a 
machine to detect. Therefore, it is harder to remove than a sharp click or 
crackle. The DH2 removes hiss by analyzing the tonal, transient and ambiance 
content of the signal at hundreds of frequency bands and removing the 
frequency bands in which it does not detect any musical signal. The operator 
must first adjust the Noise Level parameter, giving the DH2 a rough idea of 
the amount of noise present in any given signal. Next, the operator adjusts 
the Attenuation, which sets a maximum limit on the amount of noise that the 
DH2 will remove at any given frequency. Finally, the operator adjusts the 
Brightness algorithm to preserve the appropriate amount of presence by 
controlling the speed at which the DH2 will remove noise. The DH2 must also 
be adjusted by ear and is the most difficult CEDAR box to adjust. If done 
improperly, it can have a very negative result on the audio quality.
The net aural effect of removing noise from sound recordings can be 
spectacular. Judicious use of digital noise reduction can effectively free 
music from the shortcomings of its recording medium, uncovering details 
which were once masked by noise.
Listen to a short excerpt* from a record that is played first without any 
noise reduction, then processed using the CEDAR system.[...]

> (2) Why are they not perfect?
The final result of the restoration can be directly related to the state of 
the original recording.  If you have a pristine copy which was never played, 
then your work will be much easier then if you are dealing with a record 
which was played to death.  In the case of the GG recordings for the CBC, 
most were the only copies known to us (CBC and National Library of Canada's 
Gould collection) recorded on acetate discs.  Unfortunately, these types of 
discs do not age gracefully and to make matters worse, Gould's "storage" 
system wasn't what I would call proper archival storage.  Unfortunately, we 
were starting with records that were in physically very poor shape, which 
means heavy processing.  A lot has to do with the operator (see (4)).

> (3) Are there better processes for restoring original sounds?

None that I know of.

> (4) What are the main problems facing someone trying to restore a sound?

Audio restoration is ultimately done by ear, the operator programs the 
equipment to obtain the best sound with as few side effects as possible. 
 There are always trade offs when processing sound.

I find that the key to good restoration is to continually compare the 
non-processed signal with the processed signal to make sure I haven't 
overprocessed.  I try to concentrated on the sound portion of the signal and 
not the noise portion.  Concentrating on the noise portion invariably leads 
to an over-processed recording.  Something else to keep in mind is that 
there is always a psychoacoustic problem when removing high energy, high 
frequency signals.  A common initial reaction to the removal of that noise 
is that one is left with an impression that the music has lost some of its 
high frequency content but in actual fact, it really hasn't.  Overprocessing 
leaves the sound lifeless and is especially objectionable on voice.  The 
sound is muffled, as if somewhat had put a pillow on the singer.  I find 
myself gasping for air when I listen to something overprocessed.



There is a point I would like to make:  archiving (preservation) and 
restoring are completely separate issues.
"Preservation involves controlling the environment and conditions of use, 
and may include treatment in order to maintain a cultural property, as 
nearly as possible, in an unchanging state."  Unfortunately, sound 
recordings to do not remain in an unchanging state and need to be copied to 
preserved the information found on them.  At the NLC, all audio documents 
needing copying is done onto reel-to-reel tape and CDr.  Absolutely no noise 
reduction is done on the documents, everything is copied warts and all. 
 What might might be noise to one person might be information to someone 
else and what we think is the latest in technology will invariably be 
supplanted by something newer and better, and of course, removed information 
cannot be replace.

For those interested, the tape decks used in preservation work are all 
Studers and the the Ato D converter is the an Apogee AD-1000 sampling at 20 
bits and dithered to 16 using their UV16 and burned using a Studer D741.



Gilles St-Laurent
Music Division
National Library of Canada