Sonnox Restore

If you're in the market for a high‑quality restoration package, there is no shortage of contenders. Sonnox's trio of plug‑ins offers some new and unusual features and a friendly interface.

Sonnox Oxford DeClicker in Dialogue mode. Signal below the user‑defined threshold is shown in red, and treated differently from signal above the threshold.

Sonnox Restore is a suite of three problem‑solving plug‑ins designed to reduce specific types of noise artifacts in recorded material. The suite comprises Oxford DeClicker, Oxford DeBuzzer and Oxford DeNoiser, all of which must be purchased as a bundle. The software is based on a new algorithmic approach to noise removal, with a focus on minimising processing artifacts combined with an innovative graphical user interface design. Typical applications include the removal of pops, clicks, crackles, scratches, hum, buzzes, hiss and other types of unwanted background noise. RTAS, Audio Units and VST platforms are supported on both Mac and Windows PC at sample rates up to 96kHz, with protection via iLok.

All three plug‑ins use internal buffer sizes of 1024 samples for 44.1kHz and 48kHz operation and 2048 samples for 88.2kHz and 96kHz operation; Sonnox advise that driver buffer sizes be set to the same value, to optimise the efficiency of audio data transfer. As these plug‑ins are likely to be used only when recording is complete, the resulting latency times will not be an issue.

Clicks Away

The DeClicker plug‑in actually comprises three different processing sections, each with its own processing threshold control, and each designed to target noises of different durations, designated as pops, clicks and crackles. Sonnox define a pop as a noise lasting between three and 10 milliseconds and a click as anything from 0.4 to three milliseconds, while noises shorter than 0.4ms are treated by the crackle section. Each of the three sections calculates what Sonnox call an Excitation Profile: this is displayed in its own window, and is basically a measure of how far the harmonic make‑up of the sound differs from its predicted harmonic content. The idea is that a high excitation value will be detected when a large, harmonically rich event, such as a click, occurs. A click thus shows up in the excitation profile display as a triangular‑shaped peak.

The Threshold control makes the detection of excitation peaks more or less sensitive, while a separate Sensitivity control sets which of them are repaired and which are ignored. A practical strategy for dealing with clicks is to first set the Threshold just above the noise floor so that the full width of each click is detected, then adjust the Sensitivity slider so that only clicks above a desired level are dealt with. Small residual clicks can be dealt with using the de‑crackle part of the process. The display shows detected noises as 'bubbles', colour‑coded green for pops, blue for clicks and white for crackles; their size and height above the axis relates to how loud they are. If a section is disabled, its bubbles turn red, showing that they are not being corrected.

There are certain types of instrument that can produce 'false positives', such as raspy brass, so the interface includes the useful option to draw in an Exclude box to define a frequency area and level threshold window within which detected clicks should not be repaired. Only one Exclude box can be active, though you could process the same material twice with different settings if you need to. It is also possible to automate the limits of the Exclude box for different sections of an audio file, and the Exclude box can be used to prevent the unnecessary correction of small noises that are normally masked by the wanted signal.

Dialogue Mode addresses specific problems associated with removing clicks from spoken‑word material, where the process needs to be different during the speech than for the spaces between words and phrases. The obvious solution is to have two alternate parameter sets, one of which is active below a level threshold and the other above it. This is exactly what Sonnox have done, allowing you to drag a threshold line in a dynamic display of the input signal. There's also a side‑chain filter that allows you to tailor the signal used to key the dialogue/silence gate.

Operation is made easier by the ability to bypass each section and to audition the noise being removed: if you drop the threshold too far, you start to recognise wanted transients and other elements from the wanted sound being treated. For reducing the effect of vinyl crackles and suchlike, the plug‑in is excellent, and it can also deal with many short‑duration digital clicks and similar events, though there are some digital corruption issues that it can't fix completely, especially if they exceed the 10 millisecond pop duration. I'm not exactly sure how the software computes the material needed to fill the gaps, but on the whole it works really well. The Exclude feature is a nice touch too.

Buzz Off

In Auto mode, Sonnox Oxford DeBuzzer will automatically 'track' the fundamental of mains interference.

DeBuzzer is, as the name suggests, designed to remove hum and buzz — unwanted noise that is harmonically related to the fundamental pitch of the problem sound contaminant, usually resulting from mains interference. Here, the process is split into separate detection and removal sections. An FFT display helps identify the strongest harmonics, and it is possible to tune into the fundamental frequency very precisely — holding down Shift enables frequency to be accurately adjusted to a resolution of 0.0001Hz — though there is also an auto mode that follows fundamental hum/buzz that varies in frequency with time. It is also possible to treat fundamentals that vary in level.

A large, user‑adjustable dial sets the centre frequency to be targeted by the detector in the input signal, within a range that can be set to narrow, 'normal' or wide, the last producing a 5Hz window either side of the centre frequency. A frequency display shows any peaks that occur within the detection window. There's also an FFT display that correlates the audio data over time, causing consistent frequencies to stand out: the three highest peaks are marked using red lines. Finally, there's a tone generator that you can use to tune into the fundamental hum frequency until you hear beat frequencies, indicating that you are getting close to matching the pitch.

In the default Auto Mode, the detector monitors either side of the nominal fundamental frequency so that it can keep the process centred on the actual frequency rather than just the initially measured value. If the hum or buzz is fixed in pitch, you can select Freeze as soon as you hear the hum level drop. Sonnox tell us that Auto Mode is best suited to tracking a weak, slowly shifting fundamental and it works over all three frequency bands, with the option of a Fast mode for situations where mid- and high‑frequency pollution, such as lighting buzz, is varying more rapidly.

Once the fundamental frequency of the hum or buzz has been established, problems can be treated using either comb filtering — which can tackle sounds that have high‑amplitude harmonics extending over a large part of the audio spectrum — or parametric EQ, for more controlled removal of harmonics that may get quieter as they go up in frequency.

The limitation of a comb filter is that it attenuates harmonics by the same amount all the way up the audio spectrum. If you need to treat a more restricted region, the parametric EQ does a more surgical job, by allowing you to tailor the amount of cut higher up the spectrum, but at the expense of more CPU overhead. The parametric EQ can also be set to address only even or only odd harmonics if necessary, while a comb filter always takes out both. Comb filters are also less effective when the fundamental is higher than 2kHz or so. However, the notches of a comb filter are all the same width when viewed on a linear frequency scale, so each one will cover a smaller fraction of an octave at higher frequencies, and thus do less subjective harm to the audio as the frequency increases. By contrast, if you set a fundamental frequency and Q for the parametric filters, their response considered in terms of fractions of an octave might be narrower at low frequencies, but will widen as frequency increases.

There is a Range control that sets the frequency above which there are no further notches. This needs to be set so as to remove such harmonics as are audible, but without processing higher frequencies than are necessary.

Noise removal is always a compromise, and the usual strategy is to set the sensitivity as low as possible while still removing an adequate amount of buzz. An Alt mode reverses the usual sequence, so that now the harmonics are the first to appear in the audio output, followed by the fundamental, as the sensitivity is turned down. Which mode is best depends on the sonic signature of the buzz and has to be determined empirically. For broadband buzz, the default mode may be best, with Alt being better suited to hum consisting of fundamental plus only a small number of audible harmonics. Again, it is possible to audition the removed part of the signal to make sure the process isn't taking away more than it needs to.

I tested the plug‑in using an electric guitar with single‑coil pickups, held very close to the amplifier and with a high gain setting. The recording was, quite predictably, badly contaminated with hum and buzz that also varied in level and timbre as the guitar was moved relative to the amplifier. Comb filtering in Auto mode did a good job of cutting the hum without changing the tone to any significant degree, whereas with the parametric EQ I needed to adjust the bandwidth and the depth to which upper harmonics were affected to keep the tone from going slightly dull. The reduction in hum and buzz was dramatic, though if the hum level changed very rapidly it sometimes became audible for half a second or so before the software adapted to it and killed it. Considering the appallingly tough test I gave it, though, the results were nothing short of impressive.

Cutting Off Your Noise

Sonnox's broadband noise‑reduction plug‑in offers sophisticated control to help you target only the areas of the frequency spectrum where noise is a problem.

Finally, we come to Oxford DeNoiser. The broadband denoiser is always the key part of any restoration suite, targeting hiss and other unpitched continuous noise. The majority of affordable systems work by analysing a noise‑only section of the recording, which they then use to set up a threshold curve on a multi‑band filter bank, where signals falling below the threshold in any band are subject to attenuation. This is a perfectly valid approach as long as the noise is constant, though some clever band-linking may also be necessary to suppress 'tinkling' artifacts.

Sonnox's plug‑in has separate detection and removal sections for the de‑hissing process, again with an FFT display of the input signal showing the noise frequency and amplitude profile. To keep artifacts at a minimum, it is possible to remove hiss in a separate pass from more general broadband noise such as air‑conditioning noise, camera noise and so on, but what sets this process apart from simpler denoising plug‑ins is that the detector noise profile can also be set to automatic, enabling it to follow a varying noise floor. It is also possible to modify or bias the detected noise curve, and Sonnox have added a Warmth control to help optimise the subjective sound after the noise has been removed.

The Oxford DeNoiser provides different means to generate a noise profile. The default is Auto mode, where the frequency spectrum of the signal is analysed in real time to detect non‑varying frequency components, with a smoothing control adjusting response time. The computed noise threshold can be seen on the graph below the actual signal spectrum. In Auto mode, the threshold follows the signal level, so that noise removal is always be a fixed number of dB below the wanted audio. This protects low‑level information from being removed along with the noise, so as long as it produces the required degree of noise reduction, Auto operation will generally offer the most artifact‑free results. Even so, it is still usually necessary to adjust the overall threshold by ear so that the noise‑removal process isn't overdone. Colour and Air controls skew the noise‑reduction profile in the detector section to allow more or less high and/or low end to get through with less processing, making it easy to add a bit of high‑end air back into the sound.

A Freeze button can be used to fix the automatic noise profile after detection, which may be useful where a noise‑only section of the recording is available and/or the noise remains essentially constant throughout. There is also a Manual mode where the user can assume a noise profile and then make adjustments by ear. Operation is similar to Freeze mode once the noise profile has been created.

When inserted into a stereo track, the DeNoiser applies a separate process for each of the two channels, but in Auto mode a single combined left and right noise profile is derived, to preserve the stereo integrity of the signal and to avoid the possibility of odd image shifts. That means that for material where there is a radically different level of noise in the left or right channel, the Freeze or Manual modes may yield better results, but possibly at the expense of stereo integrity.

In any of the three modes, the overall noise threshold can be adjusted up or down using a 17‑segment threshold curve. A second bias curve can be used to vary the amount of noise reduction at critical frequencies, which can be extremely useful in reducing processing artifacts. The display can be set to linear, logarithmic, HF or LF views: in linear and HF, 10 adjustment handles are displayed from 2kHz to 20kHz, whereas in log or LF view, nine handles are shown from 20Hz to 4kHz. This enables you to restrict the noise reduction to areas of the spectrum where it is needed.

As broadband noise takes many forms other than circuit or tape hiss, this plug‑in has a separate de‑hiss section that is fairly aggressive but nevertheless very effective in cleaning up dialogue or other sources where there is little high‑frequency content to disrupt. You set a cutoff frequency, then adjust the reduction fader for the best compromise between hiss and dulling of the audio. The Warmth control in the output section adds back a little of the 'life' that heavy processing can sometimes remove.

There's also a Mid/Sides mode where you can use the plug‑in to clean up only the sides (difference) signal, leaving the mono or centre component untreated. This may be useful for preparing FM stereo broadcast material so that when the receiver switches to mono in areas of poor reception, the noise content remains essentially constant.

I found the broadband noise reduction fairly easy to use, having read the manual once, though you do have to be careful not to overdo the processing, especially if the original material is very dirty, as you can end up dulling it. Using the Air control or manually adjusting the threshold curve above 8kHz or so can add some life back into the sound, so it pays to get familiar with the less obvious controls. The maximum possible noise reduction is 24dB, and with normal material, you can usually reduce it by less than half that amount and still make a big improvement. The basic hiss reducer does have a significant dulling effect on sounds that contain high frequencies but it can be used to good effect on instruments such as electric guitars, which don't have a great deal of high‑frequency extension and are often quite noisy.

The Summing Up

Sonnox have managed to add both flexibility and some genuinely useful new features to the three most commonly used restoration plug‑in types, without making the process over‑complicated. As with most restoration software, the results are best when the original material isn't catastrophically noisy, but the additional adjustments allow the user to minimise processing artifacts when tackling more serious problems. The graphical user interfaces are both informative and intuitive, and I like the way you can use the plug‑ins either at a very simple level or go deeper and make more subtle adjustments if the material demands it. At nearly $2000, the Restore suite isn't a casual buy, but if you're in the business of rescuing other people's recordings and you don't have the budget to go the CEDAR route, it's well worth looking at.