Automatic removal of mouth smacks

waxcylinder, I made some improvement of progress bar behavior between my first and second versions but I know there is more I could do. If there are few or no clicks to detect at any one of the test frequencies, progress will still look erratic during the first half.

I do not fully understand how the progress bar estimates remaining time and I have opened another conversation about that in the Nyquist board.

I have not read the whole thread from years ago about Trevor’s idea of slope, clip, then integrate as a procedure for removing spikes.

My first thought is that if the slopes of the two sides of the spike are not equal in absolute value, then a dc offset is introduced, until the next spike where the offset may change. The offsets might vary in a random walk rather than accumulating. Or maybe they will accumulate if clicks tend to have sharper attack than decay, which they might, depending on the physical causes. This effect might be visible in spectrogram log f view – how does the bottom bar change?

That might explain the “pillow thumping” effect which might be mitigated with some highpass during rests, but maybe the needed cutoff would be high enough to damage low notes.

This is all armchair guessing, not experiment. Enough that I doubt the sucess.

It’s not from years ago but from October, 21st, 2013.
The DC Offset is already mentioned in the thread. There must be a DC-blocker after limiter and integrator.
This is essentially another differentiator followed by a leaky integrator.

Sorry, I must have misread member start dates for posting dates.

Some questions about Steve’s suggestions…

Would eq-band be a good choice to implement filtering of selected regions? It has hz, gain, and width parameters which can all be variable. Would it be enough to vary gain only?

In the old deesser code you shared with me elsewhere, you used a combination of high and low pass rather than eq-band. Why?

Your examples above used a raised sinusoid for the frequency envelope but a piecewise linear for crossfading. Is there any special reason one should be smooth and the other not?

Eq-band is a nice function, but I remember vaguely that the variable version doesn’t produce the same result as the fixed one (with identical constants).
Areson and Reson are also good candidates for that purpose.

I have used eq-band for similar effects with good success, but like Robert I vaguely recall there being some unexpected behaviour that make it a bit tricky to use. Sorry I don’t remember the details of that - I’ll post if it comes back to me :wink:

For very short crossfades, linear fades produce the lowest distortion, which is why they are used here.
The raised sine wave is just one option for the frequency envelope. It has the advantage that it minimises the rate of change for a given period. You may find that other methods work better - experiment.

I am experimenting with eq-band now.

I am filtering white noise and using spectrum analysis to see what the response is like. I thought it was supposed to have a broad, flat floor in the middle, but instead, with two octave width, I see a broad notchy shape. (More gradual than a “notch filter.”)

I can see from (grindef 'eq-band) that the parameters must be all constants or all sounds, and it either calls snd-biquad when they are constants, or another low-level function.

I don’t see appreciable differences yet between the results of constant parameters passed as numbers or as sounds.

How that?
Try it on a chirp (440 to 880 Hz):

(setf dur (get-duration 1))
(setf sr 8820)
(eq-band s 
   (snd-const 660 0 sr dur) 
   (snd-const -30 0 sr dur)
   (snd-const 0.5 0 sr dur))
;(eq-band s 660 -30 0.5)

Experiment with the sr parameter, You’ll get at best a single nasty click at the beginning.

Ah yes, it’s coming back to me now. :stuck_out_tongue:
For very short selections (up to about 1000 samples) a sample rate the same as the sound should work without nasty clicks.
For longer selections, control-srate should work but with one nasty click at the start if the start is non-zero.

I anticipate using this, if I do, with “short” selections of say 5 to 20 ms but that’s not under 1000 samples.

That depends on the sample rate. At the default 44100 Hz, 1000 samples is about 23 ms.

You are right, I wrote that before thinking.

I have made many changes and I think my work is at another stage worth sharing. But I will start a new thread to make sure I get people’s attention.

I followed up the suggestion to use iterations of eq-band in my experiments today. Just constant gains on each interval, and a crossfading envelope. Mostly very pleasant results, and speed is not bad.

May next try (1) broadening the fix interval a bit to each side of the identified click, (2) using variable gain, (3) overcorrecting the clicks so they are a little bit below the background.

Just for those who may be looking for some basic help, I find it easy (though tedious) to remove high-pitched noises like mouth clicks with the low-pass filter set to 6dB/octave and 1000Hz. For low-pitched noises, like when a puff of air hits the mic, try the same settings on the high-pass filter. Many wet consonants can be removed by amplifying -10dB. You’ll soon be able to distinguish clicks from wet consonants by ear.

I usually set the beginning and end of my edit to a zero crossing and do as narrow an edit as possible. Too long an edit will substantially change the quality of sound remaining, which is hard to fix.

As you gain experience doing voice acting, you can learn to hear many of your own problems and overcome some with practice. Often an immediate retake will help more than any amount of editing.

Even with much practice, it still takes me ten minutes to get a minute of clean audio. It would be VERY HELPFUL if somebody would come up with a mouth noise filter that would remove just the clicks and rumbles! :smiley:

In 2014 Paul-L made a filter for Audacity which removes clicks … Updated De-Clicker and new De-esser for speech

Indeed I did.

But you ain’t seen nothing until you have the spectral editing tools of version 2.1.0 in your kit too.

I invite jmrathbun to join the party over here. Redirecting...

Thanks! I downloaded the .ny file and put it in the addons folder so I can see it on my effects menu. In a few minutes I’ll try it out.

I applied for the suggested Facebook group and am awaiting approval.

Using the default settings, I ‘declickered’ five minutes of narration, which took about 2’15" to run. The amount of work needed for click removal was reduced about 90% with no obvious effect on audio quality otherwise. I will certainly find this add-on useful, and MUCH PRAISE for the dev!

:smiley: