Digidesign DINR

Owners of Sound Tools II digital editing systems can now have access to sophisticated techniques for reducing noise on recordings — thanks to the DINR Intelligent Noise Reduction software. Paul White loads it up and looks for a quiet life...

Digidesign's Sound Tools and Sound Tools II systems remain popular for Macintosh‑based stereo sound editing, partly because the user interface is relatively straightforward but also because the software is updated and enhanced as improvements are made. Though no software is perfect, it's probably fair to say that Sound Tools is pretty robust and does 90% of what the majority of users require from a stereo editing system.

DINR is not a part of the basic program but an add‑on noise and hum elimination module for Sound Designer II, the software behind Sound Tools and Sound Tools II. Unfortunately, the older Sound Tools system does not have the processing power to work with DINR (Digidesign Intelligent Noise Reduction); DINR can only be run with Pro Tools, Sound Tools ProMaster or Sound Tools II hardware. The host computer requirement is a Mac II series machine or a Mac Quadra, the minimum recommended memory requirement being 8Mb. The program is 32‑bit clean so it can make use of more memory if it's fitted.

Because Sound Tools is essentially a hard disk recording system, a large hard drive is essential; each minute of stereo recording uses up 10Mb of disk space. If the ProMaster hardware is being used in 24‑bit mode, an additional 50% more disk space will be required to provide the same recording time.

DINR

The DINR package comprises the DINR software, Quicktime software, and a tune‑up disk for the Macintosh System 7 (Version 1.1.1). Unlimited installation is possible from the master disk, but the master disk is necessary to complete the installation — you can't install from a copy. The installation procedure is very straightforward and clearly documented, so there's no need to go into further detail here. However, the manual I had failed to point out that the Hum Removal part of the software must run with 32‑bit addressing switched on, or it crashes as soon as you try to use it.

Once DINR is installed, it provides two new options on the Sound Designer II DSP menu: Broadband Noise Reduction and Hum Removal. These processes can be applied non‑destructively to any soundfile recorded into the system; alternatively, the soundfiles can be permanently changed. DINR can be used on either mono or stereo soundfiles in 8, 16 or 24‑bit formats at sampling rates up to 48kHz. All the essential 'controls' operate in real time, enabling you to tweak as you listen to a looped 'Preview' section of the recording held in RAM. Unfortunately, the system has insufficient processing power to allow external audio signals to be treated in real time — they must be recorded to disk first.

Because Sound Designer II can be run against SMPTE, it is quite practical to lift audio off a multitrack recorder (provided there is SMPTE on tape), process it and then re‑record it. Alternatively, complete mixes may be treated; though Digidesign make no claim that DINR rivals top‑end systems such as CEDAR and NoNoise, the results can be impressive as long as the amount of noise being removed is within reasonable limits. Notator Logic Audio users can also use DINR and Sound Designer to clean up recordings, as Logic's audio tracks are configured as Sound Designer II files.

How DINR Works

The mercifully brief instruction manual does not delve into the operating principles of the DINR system, though in reality, the Broadband Noise Reduction part of the operation is based around a digital implementation of a multiband expander; the audio spectrum is divided into 512 frequency bands, each band with its own expander threshold. The system works best if there is a small section of noise available before or after the programme material which can be analysed, but if this is not available, then the system is able to make a rough guess based on an analysis of a quiet section of the recording, the settings then being fine‑tuned by ear.

Once the material to be treated has been loaded onto disk, it is necessary to select a short section of noise‑only programme and then invoke the Learn function, which attempts to analyse the spectral signature of the noise sample. Moving on to Fit causes the software to generate a threshold curve or contour to fit the measured noise spectrum, this line being interspersed with 'grab' boxes that allow the local threshold values to be individually adjusted by the user via the mouse. The software aims to set the contour line as closely above the noise curve as possible. The entire contour line, or a selected part of it, may be moved up or down in small increments using the Mac keyboard's Alt and arrow cursor keys, and new grab handles may be added if required.

The contour line is described as an editable envelope which represents the division between wanted signal and unwanted noise. Any signals that fall above this line are treated as wanted signal; any that fall below it are regarded as noise. If it is set too high, it follows that some wanted signal will also be removed along with the noise.

If a piece of unadulterated noise is available for analysis, then the fit generated by DINR is pretty good, but if no silent passage is available (as may be the case if the recording has already been 'top and tailed'), DINR has another trick up its sleeve in the form of the Expand function. The technique involves sampling a low‑level section of the recording and then doing a noise fit on that. The initial fit will obviously be incorrect as it stands, because of the presence of the signal, but as long as the signal isn't too broadband in nature, it is possible to select just a limited part of the frequency spectrum for analysis — generally in the 10‑20kHz area, where little tends to be going on during quiet musical passages. DINR can then extrapolate this information to create a generic noise curve for the whole spectrum. Further user editing may then prove necessary by moving the breakpoints on the contour line while listening to the result.

In Practice

In order to set up the various parameters, a short section of audio may be looped in Preview mode to provide a constant source of material while the various adjustments are made. In the case of a piece of music that involves widely changing dynamics and timbres, it would be wise to check the settings on several different parts of the recording before processing the soundfile.

The DINR window is equipped with various 'buttons' and six virtual sliders. The first slider controls the noise reduction amount from 0dB to maximum; it is rarely prudent to select maximum noise reduction, as the side‑effects tend to become unacceptable. As a rule, the noisier the original material, the more likely it is that side‑effects will be audible if severe processing is carried out. An NR value of around 6dB provides a very worthwhile improvement without invoking audible side effects.

The Response slider determines how the modelling process reacts to overall changes in the noise. Depending on the type of noise, some side‑effects may be audible if the response is set too fast, but on the other hand, a very slow reaction time produces less effective noise reduction. I found a setting on the slow side of medium to be the best compromise. Release sets how fast the system recovers from changes in noise, and again, a setting on the slow side of medium is a good starting point. Artifacts to listen for when adjusting these controls include tonal dulling, noise incursion, and non‑random noise modulation which sounds like a low‑level 'tinkling'.

My first test piece was an acoustic piano recording deliberately contaminated with noise from a guitar preamp set to high gain; it is unlikely that even the sloppiest real‑life recording would require so much processing. The long decay of piano notes served to show up any undesirable artifacts quite clearly, especially when monitoring over headphones, while the bright attack of the notes soon showed up any dulling caused by over‑processing.

Though Digidesign make no claim that DINR rivals top‑end systems such as CEDAR and NoNoise, the results can be impressive as long as the amount of noise being removed is within reasonable limits.

If the noise reduction algorithm tries to work too quickly it can also generate audible artifacts, so a Smoothing slider is fitted. It is usual to start off with this in its normal position, but if side‑effects are audible, it may be moved towards the smooth end of its travel. In practice, if the level of smoothing is too high, audible dulling of the sound takes place, so it seems best to use little or none of this particular treatment.

Also in the slider department is a high shelving equaliser, designed to work only on that part of the signal which falls above the contour line — in other words, the equaliser only works on the signal, not the noise. The frequency is variable from 0Hz to half the sampling rate and the gain range extends from ‑12dB to +6dB. The idea is that any lost brightness can be compensated for using this control but without adding any noise. Similarly, it can be used to provide precise HF cut.

Once the user controls have been set, the soundfile may either be processed non‑destructively, or any section of the file selected for permanent processing. In the event that a file is to be permanently processed, it still pays to listen to it all the way through in Playback mode first, to ensure that the settings are appropriate for the entire piece. Normally there is no problem, but if the noise character of the recording changes dramatically throughout, it may be necessary either to revise the settings or process the soundfile in several sections. Processing the soundfile is time consuming, so real‑time operation is preferable where practical.

NR Contours may be saved to disk, allowing you to create and store contours for specific instruments — such as noisy synthesizer modules, effects units, or even analogue tape recorders. This also allows partly completed jobs to be resumed at a later date.

Impressions

After running my original test piece, I ran a series of passes on some remastered jazz archive material provided by Dave Lockwood, our Technical Editor. I felt that I gave this part of the system a fairly stiff test, and on the whole DINR coped very well. It is important to fine‑tune the sliders to achieve the best possible compromise between low noise and audible side‑effects, but the default settings make a good starting point. I found myself beginning with the Noise Reduction slider set to maximum, so that the Attack and Release parameters could be adjusted for minimum 'tinkling'. Then it was a matter of reducing the amount of noise reduction until just a small amount of noise became audible. This reduces the level of the tinkling sound, and allowing a little noise back into the mix also helps mask it. Headphone monitoring is highly recommended here, as it is easy to overlook some of the side‑effects when using loudspeakers. It's also sensible to keep bypassing the DINR processing, to check that the overall sound hasn't been compromised.

With the piano test recording, it was difficult to remove more than 6dB of the noise without the 'tinkling' artifacts starting to creep in, but then piano music is particularly revealing. The tinkling could also be reduced by carefully adjusting the threshold settings on the contour line, though this has to be done with care, as too high a setting can affect the natural decay of low‑level sounds.

As an experiment, the same test was repeated using a sample of a decaying piano note to provide the noise fit rather than a section of pure noise. By analysing the frequency range from 5 to 25kHz, DINR managed to generate a fit very close to that achieved using only noise. Also worthy of note is the Hi‑Shelf filter, which is very useful in enhancing the top end of a recording without adding noise. Alternatively, the filter may be used to remove high frequency noise from programme material that has a limited audio bandwidth.

Summary

DINR's Broadband Noise Reduction module can reduce electrical hiss, mechanical noise (such as that produced by air‑conditioning) and low‑level buzz from lighting systems, though the higher the degree of noise contamination, the more audible the side‑effects. Pop music mixes fare pretty well, as they tend to maintain a high average energy all the way through, which helps mask any existing noise and means that a lower level of processing may prove sufficient. The same is true of individual electric guitar tracks, and if the recording can be processed before sound effects are added, it is unlikely that any untoward artifacts will be audible in the final mix.

More delicate material can be treated with the proviso that the degree of noise reduction applied is moderate and that the user is prepared to spend a little time optimising the control settings. The only time the system struggles is when the noise level or character varies widely throughout the programme, as might be the case when the recording contains several edits or where mix automation has been used to mute or change the level of noisy sources during the mix. It is quite possible to treat different sections of the recording with different DINR settings, but most often it is only necessary to ensure that quieter passages are processed sympathetically; it's usually OK to leave the louder ones to look after themselves, as any changes in noise level will be concealed by the natural masking effect of the material itself.

I felt the Hum Removal filters were a little disappointing, and in my opinion they don't work as effortlessly as those in something like Roland's hardware‑based SN550 single‑ended noise reduction system. Unless the filtering is kept to a fairly gentle level, they can be heard working, and I didn't feel they made a particularly good job of targeting just the hum. Nevertheless, when used in conjunction with the Broadband Noise Reduction module, the improvements that can be made are still worthwhile. On balance, I feel DINR is worthwhile for its broadband noise reduction capabilities alone, so perhaps the failings of its hum filters can be forgiven.

Where noise reduction is concerned, there's still no such thing as a 'free lunch', but for existing Sound Tools users running with Pro Tools, Sound Tools ProMaster or Sound Tools II hardware, an affordable DINR is the next best thing!