Audio Restoration Workflow

Introduction

Back at the birth of digital audio restoration in the late 1980s, there was a defined order in which problems were tackled. First, you removed the clicks, then the crackle, and then the broadband hiss. There was a simple and compelling reason for this. The spectral subtractive dehissers of the day were blunt tools when compared to 21 st century noise reduction and, if you applied them first, they would remove some of the information that allowed the declickers and decracklers to identify the problems they were designed to remove. So the rule was declick – decrackle – dehiss, and it never varied.

Things became more complicated when new processes were introduced. Should you debuzz the material before declicking it? When is the best time to use speed correction? Should you use spectral editors to remove individual noises before or after dehissing? And so on…

It’s sometimes useful to consider the order in which noises and other errors are introduced into recorded audio. To illustrate this, let’s return to the dawn of the recording industry and consider a performance from the early 20 th century. The band would play in front of a recording horn (or later a microphone) and the audio would be cut into a wax cylinder or disc. If the medium was a 78rpm disc, this would then be used to create a stamper that would itself be used to press commercial discs. Many years later, as these media decayed, the audio would be transcribed onto analogue tape. A few decades after that, the audio on the analogue tape would be recorded onto a digital medium. Following this, the data comprising the audio might be compressed to store it more efficiently, or to transmit it on the low-bandwidth digital networks of the era.

At every stage of the process, problems could be introduced. Even if the recording was made in a quiet and well-treated environment, the recorded sound might have significant resonant peaks and troughs that rendered its tone quite unlike that of the original performance. Making a commercial disc would introduce further issues because dust would create deviations in the groove walls, and the coarse materials used for commercial discs introduced further departures from the ideal. Playing these discs then added further problems as the grooves were worn away, and careless handling introduced physical scratches in the record surface.

Further problems were added when recordings were transcribed to analogue tape. Even with suitable biasing, tape is hissy, so a layer of broadband noise would be added. Furthermore, there’s no guarantee that the tape heads were adjusted correctly, and mistakes here would add azimuth errors that led to timing errors between channels and smearing of the signal in each channel. Analogue tape machines are also prone to speed inconsistencies, leading to low frequency deviations (wow), higher frequency deviations (flutter), and other aberrations such as slowing down or speeding up as the tape is recorded or replayed. Later, these problems are ‘baked in’ when the audio is transferred to a digital medium.

Unfortunately, this isn’t the end of the problem creation, because digital recording and playback systems can also be prone to dropouts and clicks caused by dropped samples and clocking errors. Next, there are the degradations introduced by discarding large amounts of audio data to create MP3s and other common formats. And, finally, when you hear the recording, you may discover that somebody slammed a door in the studio just as the band and singer were performing the quietest and most emotive part of the song. It’s little wonder that audio engineers struggled to release vintage material on new media before the advent of digital audio restoration. It could be done, but the people with the skills to rescue these recordings were few and far between.

Peeling the onion?

Today, there are all manner of affordable tools promising to remove each of these problems but, before going any further, let’s be very clear about something. Removing noise is simple. You can do it with complete success by turning the volume of your playback system to zero. You might feel that this defeats the purpose, and you would be right; you have removed all of the problems, but with an unacceptable side effect – you have also removed all of the wanted music. This illustrates an important point: removing noise is easy, but doing so without introducing unwanted artefacts is hard. How many times have you heard a processed recording that sounds lifeless or has been infected by twittering or glugging sounds? All of these can be the side-effects of unsympathetic audio restoration. The golden rule here is to concentrate on the quality of the wanted signal in the output rather than concentrating solely on how much noise you can remove.

So now we return to the point of this paper. How can we use the restoration tools at our disposal to rescue the audio? An important part of the answer lies in applying those processes in the right order.

It’s tempting to think that we can restore damaged recordings by removing the problems in the order in which they were inflicted but, while this works in some cases, it’s not always correct. The most obvious example is where a scratched record has been recorded to analogue tape, adding a layer of hiss over the audio. If you reduce this hiss first, you may not be able to remove all of the impulse noises: clicks, ticks, pops and so on. So here’s a guideline to the order of processes. It’s far from all-inclusive, and there will be numerous cases where applying them in a different way will achieve superior results, but it should provide a good starting point for much of your material.

Declipping

Clipping occurs when something in the signal chain is unable to handle the signal levels presented to it. This is often revealed as a series of flat tops and bottoms visible in a suitable waveform display. If these are present, use a declipper to remove the flat sections and regenerate the signal that would have existed before the clipping occurred. It’s vital that the audio level is reduced at this point or the reintroduced peaks will be clipped again. Note that you will be unable to correct a clipped voice or instrument within an otherwise unclipped mix.

Declicking and decrackling

These processes should be applied before anything that changes the signal's bandwidth (such as variable-rate sample rate converters in azimuth correctors and speed correctors). Furthermore, many of the processes that follow are performed in the frequency domain, which can degrade the information used by declickers and decracklers. Always declick first, removing the obvious clicks and ticks. Don’t be tempted to over-process to try to remove underlying crackle using the declicker. This may not be effective and could lead to artefacts. Remove the tiny impulses that comprise crackle (which are often not visible on a waveform display) using a dedicated decrackler.

DC filtering

View the audio waveforms in a suitable display. If a DC offset has been applied to any of the channels, you’ll notice that the bulk of the signal lies above or below the centre line. Correct this using a dedicated DC (high-pass) filter.

Manual declicking and dethumping

If the audio suffers from extended clicks and thumps, it’s possible that these will not have been removed by a streaming declicker. If you attempt to push the declicker harder to address these, you will introduce artefacts, so you should remove them using a manual declicker and a dedicated thump remover. Take care to retain the atmos in the corrected regions. If you don’t, you will obtain holes in the ambient noise where the extended clicks and thumps previously existed.

Debuzzing and speed correction

The next thing to check is whether the pitch of the recording is consistent. If a buzz is present, you must decide whether this was imposed after the speed variation occurred (in which case it should have a constant pitch and must be removed before speed correction) or whether it lies within the original signal (and should therefore be removed after speed correction).

Constant tones unrelated to the music or speech in the recording are often (but not always) generated by 50Hz or 60Hz mains signals bleeding into the audio chain during the various recording and transcription processes. If the pitch of a buzz or hum is consistent, employ a debuzzer at this point. Don’t attempt to eliminate the fundamental and its harmonics using EQs, no matter how tight the notches. This will lead to a hollow sound that cannot be later corrected.

Following the removal of steady buzzes and hums, you can apply speed correction. You should find that any buzz or hum present in the original recording now has a constant pitch, and you can remove it using your debuzzer.

Azimuth Correction

If your material is stereo or multi-track, you should now check whether the channels are synchronous. If not, use a phase (or azimuth) corrector to line them up. In general, you can’t do this by sliding one channel against another in a DAW because the errors will not be quantised to exact samples, and the offset may be fluctuating throughout the audio. Always perform azimuth correction after removing buzzes or hums, otherwise the detector may lock to these and fail to correct the error in the wanted signal.

Retouching

Often, the penultimate process is spectral editing or ‘retouching’ to remove discrete noises such as coughs, dropped items and slamming doors, or extended problems such as wind noise. The key here is to identify each unwanted sound on both the time and frequency axes of a spectrogram, treating it without affecting other sounds that exist in the same region of audio. As much as possible, ensure that the background noise remains (or becomes) consistent. This will give your broadband noise reduction processes the best chance of removing the remaining noise without introducing artefacts.

Broadband noise reduction

There are many noise reduction processes with names such as dehissers, noise suppressors, and voice extractors. In general, these should be used as the last stage in the restoration chain. This is where you are most likely to create artefacts, so be careful to use an appropriate algorithm and listen to the output to ensure that none are being introduced.

Audio Sweetening

Once you have the cleanest signal you can obtain without introducing artefacts, you may wish to sweeten it using equalisation or matching EQ, and all manner of processes to mould the desired dynamics and overall loudness for the final output. The decisions you make here are subjective and choices should be applied with care to obtain the characteristics wanted for each given project. For example, a slight boost of the high frequencies coupled with gentle limiting may help to reveal detail in a vintage or low-quality recording, but a heavy-handed approach will lead to a harsh result that becomes tiring with extended listening.