Which Audacity clean-up effects should be distortion-free?

[R_G_B]

In context of recording to Audacity from vinyl or tapes, I am wondering which of the common 'clean-up' effects are expected not to add distortion.

I would be relaxed about 'routine use' of those unlikely to add distortion. For example I am thinking about routine use of a 20 Hz high pass filter to remove DC offsett as discussed here: http://forum.audacityteam.org/viewtopic ... 28&t=10216). But I would think longer before applying effects that likely add distortion (and favour another solution like a more expensive sound card to eliminate minor noise or offsett; or more time spent tweaking the recording to just below clipping on the highest peak rather than using effect > amplify later if amplify also distorts).

The manual tells us that Noise removal may have a distortion penalty. (When I use it on 1.3.12-beta under Win 7, I see no change after clicking Preview, but a big change after clicking OK.)
Notch requires human help, so an excessive setting is likely to be noticed on the spectrum (http://wiki.audacityteam.org/index.php? ... tch_Filter).
But I could not find a comment about other common clean-up effects.

Any advice from the experts about whether to anticipate added distortion from the following at default settings?:

* 20 Hz high pass filter

* 20KHz low pass filter (probably a waste of time with 44100 Hz sampling)

* Click Removal

* Normalize vs Amplify (assuming balanced stereo tracks and no DC offsett)

[steve]

* 20 Hz high pass filter

Beneficial for removing sub-sonic noise. No affect on the "audio" range (20 Hz to 20 kHz)

* 20KHz low pass filter (probably a waste of time with 44100 Hz sampling)

No benefit for 44100 Hz sample rate. No affect on the "audio" range (20 Hz to 20 kHz)

* Click Removal

Can benefit tracks where clicks are present. Should have minimal impact on normal audio, though there may be slight affect on some transients (may slightly soften sudden percussive sounds) so for critical audio it can be worth testing on short sections before commiting.

* Normalize vs Amplify (assuming balanced stereo tracks and no DC offsett)

No distortion, though Normalize will change the stereo balance if the maximum peak in one channel is greater than the maximum peak in the other.


The above assumes that processing is done on 32 bit tracks. All effects will produce some (though usually extremely small) distortion with 16 bit tracks due to the limited precision of 16 bit.

Exporting a 32 bit track as a 16 bit file will add a tiny bit of noise due to "dithering". This is usually preferable to "rounding errors" from down-sampling without dither. http://en.wikipedia.org/wiki/Dither

[R_G_B]

Thanks stevethefiddle, that is very helpful. The link gives me new patience with ditherers.

In theory I guess it should be better to get the original record level right, so there is only one cycle of rounding / dithering in the initial digitization; whereas effect > amplify imposes a second cycle of rounding / dithering on every sample that does not fall exactly onto one of the available digital levels after the amplification? In Audacity does '32-bit float' mean that all of the effect calculations are done at 32-bit (4,294,967,295 levels) even if the sound card supplies signal only at 16-bit (65,536 levels) or 24-bit (16,777,216 levels)? In that case I can see why effect > amplify calculations (and other simple subtraction or scaling) need not add perceptible error above the limits of the initial digitization and export into 16-bit (let alone MP3).

One thing I noticed is that use of the 20 Hz high pass filter to remove DC offsett causes a few regions that were just below 0 dBFS to become clipped.
I know, that is why you advise to "record at a lower level to allow some more headroom for the calculations".

[kozikowski]

Audacity doesn't use 32-bit. It uses 32 floating which allows for magic like undistorted sound level over "0" during effects.

We determined through exhaustive posting that the DC Offset tool in "Normalize" worked better than trying to do it by hand or with the equalizer. It takes the average of all the samples rather than trying to separate 19Hz from 21Hz. Theory has it that you can have sound that makes the Overall Average fail, but nobody in 24 time zones could come up with a real world example of that failure.

If you have fractional dB sound management causing clipping errors, you are totally capturing the material too hot. Capturing hot can mess up some of the tools, too. If no waveform ever gets seriously hotter than the show average, the clip fix and some of the noise reduction tools will fail.

Koz

[R_G_B]

Thanks Koz,

From your advice in http://forum.audacityteam.org/viewtopic ... 28&t=10216 I gathered that frequencies below 20 Hz should always be removed, which also removes the common source of DC offset; so I figured why apply one more effect (with whatever small added errors) through the option in Normalise? Seems I jumped to the wrong conclusion.

The maths are out of my league, but if I have it straight: '32-bit float' means effectively a +/- 24-bit scale for the mantissa, plus an 8-bit exponent. This sacrifices some of the precision of '32-bit integer' for much wider range - which not only assists mathematical programming but ensures that intermediate results during the calculations do not go off scale. I gather that the end result is still that the 'rounding errors' within the 32-bit float processes should be imperceptible relative to those in the initial 16- or 24-bit (integer) digitization by the sound card, or the final down-sampling for export into 16-bit (integer) WAV or any MP3. Both human hearing and electronic noise in DACs have practical limits below 140 dB dynamic range, which is less than the number of levels that can be encoded in 24-bit (more than can be encoded in 16-bit but I gather it would need an amazing recording, a professional equipment chain with no other weak links, and a good pair of ears to hear the difference).

Some other Effects (not those listed above) involve lots of other compromises that are more likely to introduce distortion for reasons other than 'rounding errors' in the calculations.

(PS: I now understand why Preview shows no effect under Noise Reduction. It is actually Pre-listen on a short sample, not Pre-view on the waveform display. So the 'Preview' is nothing {silence} if Noise Reduction works on a quiet section sampled. A trick for novices to discover with a chuckle - in my case after Pre'view'ing a different effect.)

[kozikowski]

<<<a professional equipment chain with no other weak links, and a good pair of ears to hear the difference>>>

Effectively impossible because straight electronic components and electrical connections have noise and can be relied on to limit any design project. My audio design books feature a chapter on when to give up.

"Mount the amplifier above a chamber of liquid nitrogen..."

Koz

[steve]

Layman's terms: 32 bit float = accurate enough to be considered perfect (and fast processing).

In theory I guess it should be better to get the original record level right, so there is only one cycle of rounding / dithering in the initial digitization; whereas effect > amplify imposes a second cycle of rounding / dithering on every sample that does not fall exactly onto one of the available digital levels after the amplification?

Better to allow a bit of headroom so as to avoid any possibility of clipping, and set Audacity to record at 32 bit quality. Using a 16 bit sound card (or a higher resolution sound card with 16 bit drivers), the recorded audio will still only be 16 bit, but converted losslessly to 32 bit (there will always be an exact 32 bit float value for each 16 bit sample). However, because you now have the track in a 32 bit format you can amplify up and down as often as you like without introducing any significant losses. The processing is done at 32 bit and the track is 32 bit, so no format conversions and no dithering. The down-sampling and dither will only come into play when you export your finished project as 16 bit.

One slight caution - although all processing is done with 32 bit float, some effects may still clip at 0 dB. Most effects don't, but I think there are still some effects that do, so the recommendation is to use "best practice" and keep your levels below 0 dB.

One thing I noticed is that use of the 20 Hz high pass filter to remove DC offsett causes a few regions that were just below 0 dBFS to become clipped.

You should be able to get away with that in Audacity 1.3.12 - it has probably not actually clipped, but just gone over 0 dB. If you apply the Amplify effect it will probably default to a small negative amplification amount. That amount is the degree of attenuation required to bring signal back down to a 0 dB range.

As a general guide, allow about -6 dB head room while recording. Set Audacity "Quality" settings to 32 bit float (in "Edit menu > Preferences > Quality"), and during processing, amplify the track as necessary so that whatever effects you use do not cause the waveform to exceed 0 dB. Immediately before Exporting your finished masterpiece, Amplify with a target level just below 0 dB (say around -0.3 or -1 dB). For high quality MP3 exports it is best to keep the peak level no higher than -1 dB as the encoding process can sometimes cause peaks to go higher than the original uncompressed audio.

[R_G_B]

Many thanks Stevethefiddle and Kozikowski, I have learned much more than I set out to ask.

You should be able to get away with that in Audacity 1.3.12 - it has probably not actually clipped, but just gone over 0 dB. If you apply the Amplify effect it will probably default to a small negative amplification amount. That amount is the degree of attenuation required to bring signal back down to a 0 dB range

I tested this and you are exactly right Steve. Thanks also for the MP3 export level tip, I had no idea to use Amplify in those ways.

Koz, by coincidence I use liquid nitrogen to reduce electronic noise in another application (a cooled, back-illuminated, laser-thinned, slow-scan CCD for ultra-low light imaging). Not planning to try it on my sound card. However, there is that spare iMic in your garage ......

Much obliged, RGB

[kozikowski]

<<<However, there is that spare iMic in your garage>>>

LN2 will not cure bad design, unless you rap the device sharply after exposure.

Koz

[waxcylinder]

I use Audacity for most of the production chain - but I really recommend using ClickRepair for click and pop removal. It costs about $40 but is well worth it for the time and effort that it saves - and I find the results just a little shy of magic. See this sticky thread http://forum.audacityteam.org/viewtopic.php?f=28&t=1994

WC