Metablog

Thanks for this fantastic compressor. I’m using it to level the speech in the latest episode of my podcast. The one built into Audacity must have a non-zero attack time, which results in every new segment of speech beginning with a loud blast. I was actually looking for Nyquist samples so that I could write my own leveler. You’ve not only saved me hours of fine-tuning per episode, but days or weeks of research and development.

I record each person on a separate track, so there are long periods of silence on each track while that person is listening. I’m experiencing a strange artifact during those portions where the volume would turn up and I would hear that person breathing. But I think that adjusting the floor will take care of that problem.

Thanks for all your hard work. It’s a real time saver.

The program is excellent - but it crashes when I try to enhance long (90 minute) files. The files are 24 bit or 32 bit floating, and I’ve tried 1.3.2 and 1.2.6. It takes a long time for the crash to happen, but it does, perhaps a memory leak or something. Enhancing sections of the same files (1-5 minutes or so) has worked perfectly. I haven’t found the threashold yet…

It’s probably not a leak, just using too much memory. Try changing this line:

(defun get-my-sound (sound)
(snd-avg (if (= use-percep 1) (get-percep-adjusted-sound sound) sound) 2000 1000 op-peak))

Change the numbers to 10000 5000, and if that doesn’t work, to 20000 and 10000.

That makes it so you get a less fine-grained view of the source sound, so if you’re using relatively fast attacks or releases, you might notice that the volume changes slightly too soon or late the larger those numbers get. You can lower the larger number to closer to the smaller, and that will increase the chance of clipping but make the volume changes slightly better timed.

You might be able to apply it in sequence without introducing clicks by making the edges of your selections be high peaks in the file.

Wow! This is exceedingly convenient. I, for one, appreciate this very very much.

Will try it - thanks! Also, is it possible to set the attack and release to faster than 20db/sec? I have gun shots in a live sound track over voices, and want to avoid hard clipping.

-David

For those, I would definitely try to manually correct for the individual gunshots. (The Nutcracker?) Audacity has a volume envelope tool that you can use to precisely deal with that sort of thing. If you still want to use the plugin to do that, at the very top of the .ny file there’s the definitions for the parameters, that you can edit to change the allowed range.

Wow, your compressor works really well. I agree with Michael, it’s good enough to use for podcast and (homebrew) radio production.

Hope you don’t mind a feature request:
Is it possible to add a flag, so that the compression would only be performed on a single channel (left/right) based on either the volume of this channel or the combined volume of both?

This would allow mono “talk-over” effects if you record your spoken track on one channel and have music on the other. First you could compress the voice channel and then the music in a way that it whenever something is spoken on the other channel, the music stays in the background to immediately return afterwards.

Dan, I think that the option you’re looking for is really more of a separate plugin. The basic idea would be to take a volume envelope of the compressed voice track, invert it, and scale the background track to follow the inverted envelope. That sounds like a five line plugin to me, but I’m not up to writing it anytime soon. If I had to guess, it’d look something like

(mult background-snd (invert (snd-avg voice-snd 2000 1000 op-peak)))

where “invert” isn’t actually a Nyquist function, and you extract the “background-snd” and “voice-snd” sounds from the appropriate places.

You could try putting a request to the audacity-nyquist list, though.

Can someone help out a newbie with some suggested settings for the parameters? I am using this on a conference call recording that has all sorts of people talking with lots of different amplitudes.

You might want to increase the rise and fall speeds to 10 db/sec, and raise up the noise floor to .3 or higher.

Hello, is it really Chopin’s Ballade no. 1 in G minor? The one I have at home does not have that part in it at least. Just curious :-)

Take care.

Oh my. I had Chopin’s ballade in the sound clips at one point, but then I substituted them for Debussy’s The Snow is Falling, from the Children’s Corner suite. I guess I forgot to change the text.

[...] Metablog » Chris’s dynamic compressor (tags: compressor compression audio audacity nyquist plugins opensource freeware mac linux osx windows winxp) [...]

Great plug-in! It’s weird no one has posted since last October. Is there a way to make the settings more precise, so I can adjust the floor in increments of .001? Thanks! —Allen

Heh. I think I had some overzealous spam protection there for a while.

I don’t directly control how precisely you can set that parameter. If you just type in the values into the text box I bet it will let you be as precise as you wish, up to the limits of double-precision floating point.

Incidentally, I’m about to release a new version with a new, better algorithm. I have initially written it in C# but could make it work (somewhat more slowly) in Audacity. If you have a standalone tool, would you still want the audacity plug-in? (If I don’t write a new audacity plug-in I’ll leave the old plug-in.)

Thanks! I had played around with some very small values like .002 and convinced myself that they were getting rounded to zero, partly because I couldn’t hear a difference and partly because when I re-opened the effect dialogue after, that field said zero again even though usually the last values used are remembered. But neither of those is conclusive. (And as for values being remembered, there seems to be a difference in other Audacity effects as well between typing them in vs. using the sliders… sometimes it seems like only using the sliders works, but I haven’t tested carefully yet.)

I’ll look forward to your new version. From my perspective, I think an Audacity plug-in would be useful. To use a standalone tool I assume I’d need to export the audacity project to a single file and then re-import, right? Plus I use a Mac… is my sense correct that it’s hard to make C# programs run on Mac?

Thanks again,
Allen

Huh. Well, if you’re seeing odd behavior then you could always change the plugin source so that the range of the noise floor is smaller. That way the slider will be more detailed. It’s the 10th or so line:

;control floor “Floor” real “linear” .02 0.0 1.0

Just lower that 1.0 to .25 or some such. Be warned that any changes to the .ny file will cause Audacity to forget the saved values of all the parameters and go back to the built-in defaults.

The Mono framework should allow programs such as mine to run on a Mac. It requires the 1.9 version, though, which is still prerelease and a quite large download—95 MB or so.

Hi Chris,

I’m still getting lots of milage from your compressor. Thanks for the advice about changing the slider ranges; that lets me know for sure my settings are what I want them to be. I have another question about the noise floor, though — I’ve finally realized that it doesn’t behave as I had assumed. Sections of audio beneath the noise floor are still raised, just not quite as much as if you don’t set a noise floor. So if someone doesn’t speak for a couple seconds, you get a little hill of background noise in the middle of the break in speech. So I changed “(defun limit-amp (samp) (if (and (> floor 0) (> samp max-gain)) max-gain samp))” to “(defun limit-amp (samp) (if (and (> floor 0) (> samp max-gain)) 1 samp))”. This sort of does more like what I’d expect, though of course if something just barely breaks the noise floor, up it comes. But I barely read Lisp, so I still don’t know how most of the compressor works, and I thought I’d get your reaction to this, and see if you had any other thoughts.

Thanks,
Allen

Yeah, I think what you’re looking for is a “noise gate”. Lots of compressors have them built in, but I don’t have one in mine because it’s not really appropriate for most classical music recordings, which is what I wrote the plugin for. Alas, it seems that most people use it for speech instead. I’m not sure how I’d go about implementing a good noise gate algorithm, because I haven’t ever thought about it. The naive noise gate algorithm is just what your change consists of, and I’m not sure how much you can improve on that. But you might look for another plugin to apply in a separate pass (probably before compression).

The other thing I might do is decrease the compression ratio and attack/falloff speeds a bit. (Normalizing before compressing would help too.) That way the compression isn’t much harsher overall, but noise isn’t boosted as much.

I guess a more general solution would be to generalize the compression ratio parameter into a function with a domain of input dB and range of output dB, which is how SoundForge’s compressor works. It would probably be relatively easy to implement, but Audacity doesn’t offer a good GUI for that, or a way to create one, so you’d have to tweak the file itself every time you changed the function. The function would just replace “adjust-samp”, and look something like this, I guess:

(defun adjust-samp (db)
  (db-to-linear
    (cond
      ((< db -30) 0)
      (t db))))

“cond” is like a switch statement, with “t” being the default clause. The maximum amplitude and noise floor parameters would become superfluous. The sample function there silences everything below -30 dB and applies no change to everything else. But ideally if you graphed the “adjust-samp” function it would be continuous. I don’t think adjust-samp actually works on samples, btw.

Thanks, Chris! This is helpful. I will play with noise gates, and with the adjust-samp function once I learn enough Nyquist to understand what’s going on. So do you think we’re crazy to use your compressor on speech? The conventional compressors may have some advantages, but the thing I like about yours is that it doesn’t miss any spikes. The Audacity compressor misses spikes (like mic hits or cassette tape artifacts) during compression, but then the spikes keep the normalizer from bringing up the volume, so I end up having to remove the spikes by hand. (Click removal doesn’t really seem to work; I’m not sure why.) Also, what do you mean when you say you don’t think adjust-samp works on samples?

Allen

“So do you think we’re crazy to use your compressor on speech?”

Hey, if it works for you that’s great. I’m not surprised it does better than other sorts of compressors for some things. Nyquist comes with a compressor that’s similar to mine, but Audacity’s repackaging of Nyquist doesn’t include it (or a couple of library functions it uses).

“Click removal doesn’t really seem to work”

Well, clicks, pops, and other artifacts may sound similar but methods used to detect them can vary widely. For a microphone hitting something, I usually do a high-pass filter at around 20-50 Hz. It doesn’t get rid of the sound, but it gets rid of the peak in the waveform. For speech, I’d probably do a highpass to the whole thing. For dealing manually with big peaks, it’s pretty easy to use Audacity’s interactive envelope function to scale it way down.

“what do you mean when you say you don’t think adjust-samp works on samples”

The value passed to adjust-samp isn’t an individual sample, but part of the envelope that gets interpolated over the waveform. adjust-samp gets called once for each thousand samples.