De-clicking Noisy Transfers (2)


Last August I started evaluating a software utility called ClickRepair (authored by Brian Davies) which processes digital sound files to reduce audible clicks and crackle. Software of this type is particularly useful for transfers made from 78s because such noises are inescapable with the shellac material from which the source discs were originally made.

I did lots of experiments and came to the conclusion that a considerable proportion of my own music library would benefit from processing. Ironically, it is the ‘best’ transfers that benefit the most. A lot of my CD-sourced music files are ‘digitally remastered’ from LP compilations of original 78s. Frequently, that is a euphemism for them having been equalised, compressed and filtered almost to death, with little or no high frequency component and often with ghastly added reverb or ‘stereo’ effects.

More recently I have acquired a significant number of modern transfers (nearly all by TangoTunes) that have aimed to extract the maximum possible fidelity direct from the source material. Shockingly, the main recording companies deliberately destroyed all their masters in the early 60s, so we now have to rely on commercial pressings which are very old: original 78s from the early 1930s onwards – and they are frequently worn, scratched and crackly. A faithful transcription of such material is inevitably going to extract a good deal of noise along with the music, but because the two are completely interwoven, the noise comes along for the ride.

However, clever software can ‘examine’ the audio signal, and can identify clicks and crackle. There are several such applications available, and they basically all work by splitting the music file into many thousands of very tiny parts. Where a click is identified, that part of the audio signal is repaired, leaving the remainder of the file untouched.

D’Arienzo: El choclo (1937)

Here’s a short extract from the TangoTunes transfer (taken from the album ‘todo de juan (v3) D’Arienzo (1935-39)’):

The graph shows the frequency range present in the sample: impressive for 1937.

Next, here’s the same extract, processed with ClickRepair, using the ‘default mono’ settings: that is de-clicking at 50% sensitivity, with de-crackle off:

As most of the noise is crackle and hiss (rather than clicks), the sound isn’t dramatically different. The application splits the 20s extract into 882,000 samples, and clicks were ‘repaired’ in 1,786 samples (1/494) – a very low rate of intervention.

The next sample was processed using the ‘default 78’ settings: de-clicking at 50% and de-crackle at 50%. This time 13,146 samples were ‘repaired’ (1/67):

Finally, the last sample was processed using custom settings: de-click at 50%, but de-crackle at 90%. The ‘repair’ rate is now 52,224 (1/17):

My own subjective listening tests tells me that if the repair rate is around 1/100 (that is, that 99% of the file is completely untouched, and 1% is processed), I generally prefer the processed version to the original and cannot hear any noticeable detrimental effect on the music. Where the source file is very noisy, I sometimes prefer a higher level of processing, but that about 1/50 is the limit. The third sample, here, was processed at a much higher level – but subjectively, the result is still acceptable to my ears. Nearly all the crackle has gone, leaving just a steady hiss. Other processing tools could be used to reduce that hiss too, but it doesn’t trouble me, and reducing it does real damage to the integrity of the remaining music signal.

For completeness, here is a final sample that has no music, but just the ‘noise’ that was extracted from the three processed samples (6s each). You can clearly hear the noise level go up three times. The third sample has the most aggressive processing: would you really want to add this back in to the music? Listen, again, to the original and decide.

The six D’Arienzo albums that included El choclo were the first in TangoTunes’ ‘Golden Ear’ series – and they were generally very good indeed. Some of the source material, released as early as 1935 is significantly compromised to begin with, but these are the best transfers of this repertoire I have heard.

Pugliese: La abandoné y no sabía (1944)

Here’s a short extract from one of TangoTunes’ most recent releases (taken from the album ‘San Pugliese (v1) Pugliese (1943-45)’). First published in December 2015, this is from the revised edition (February 2017), which applied different values to the critically important de-emphasis process. These are among the best shellac transfers I have ever heard. Odeon were achieving marvellous fidelity in their original direct-to-disc recordings by 1944 – tape mastering was still almost a decade away. This is just the sort of repertoire that benefits hugely from taming the running noise from a shellac disc that is probably 70+ years old.

I chose this song to illustrate music with a very quiet opening and a wide dynamic range. The unprocessed extract first:

The next sample was processed using the ‘default 78’ settings: de-clicking at 50% and de-crackle at 50%. This time the ‘repair’ rate was 18,242 out of 882,000 samples (1/48):

And finally, the last sample was processed using custom settings: de-click at 50%, but de-crackle at 90%. The ‘repair’ rate is now 51,372 (1/17):

Despite the high level of processing, the application has done a good job of rendering the music with no more than steady hiss. This is a good example of why the best transfers benefit from processing: the clarity of the resulting sound is remarkable.

By choosing to process these files, myself, I imply no criticism of TangoTunes for publishing them in a more raw form. Some will hear damage done to the integrity of the sound through such processing. We are free to choose. My own preference (and that’s all it is) is to tame the running noise and enjoy the music more.

De Angelis: Cristal y luna (1948)

My last example is taken from a Reliquis CD (20 Exitos v1). The frequency and dynamic range of the transfer is no equal to the TangoTunes samples already considered, but the original is quite noisy:

Just one processed variant, this time: using the ‘default 78’ settings with a repair rate of 1/65. The result is quieter, but ironically, the reduction in the noise makes the limitations in the sound very apparent. I still prefer it, but the sound is not very good.

Being able to automate the removal of clicks, at all, seems a very neat trick, but if you look at this sample waveform (from the De Angelis file), two clicks are quite obvious. Crackle is harder to ‘see’, but the application generally does a good job of finding it.

It has taken many, many hours to go through my library and decide which files to process. My working method has been to try the ‘default 78’ settings first on whole albums and then to review the results. The software allows batch processing and a log file can tell you at a glance what the ‘repair’ rate was, file by file. Higher intervention suggests noisier source material, and listening tests confirm whether different settings would render a more acceptable result.