Here are the latest improvements to the plug-in I am developing for filtering out of wet mouth noises and clicks from my narrations. I also present a de-esser using similar methods that fell out as a by-product.
De-Clicker is designed for filtering of natural but undesirable noises that take a few milliseconds to decay, not for repairing spikey damage to a digitized signal. I did not intend this for music or vinyl repair, but I would like to hear whether it has unexpected other uses.
The default settings in both tools are the ones I favor now for treatment of whole tracks. Do you have different opinions about good default settings? Let me know.
I successfully treat most of the more resistant clicks by hand with a single pass and double the number of bands: NOT by lowering the dB threshold control. But such settings may be too slow and too aggressive for batch treatment. Yet the nice thing is that you do not have to select the interval around the click very narrowly. The surrounding signal is usually unaffected. So select it and fix it and move on, donāt fuss with zooming in and out.
I treat typically no more than a half hour at once. Compute time does unfortunately go nonlinearly with length of the track to some extent. Not sure why. (Iāll blame garbage collection.) But I have improved performance much over the previous version.
As mentioned below, do low-frequency rolloff (highpass filtering) before De-Clicker for slight improvement of some results.
Detailed advice about settings:
In either tool, choose Apply Changes to hear just the results. Or, choose Isolate Changes to hear only what is subtracted from your signal. You donāt want to hear a lot of āmurmurā so that you almost understand the words: your settings are taking overtones out of your vowels and muffling the voice. You may notice that the attack of many consonants is somewhat affected, but you might judge that this is acceptable and sometimes even helpful.
If you duplicate a track, then isolate changes in one, you can listen to the tracks together to hear the repaired sound (the isolated clicks interfere destructively), or mute one to hear just the original or just the subtractions. You might silence corrections you donāt want before mixing. I donāt do any of this once I find settings I trust.
Both tools specify a range of frequencies and a number of frequency bands (of equal width in log-frequency) for detection and repair. Impatient for results? Use fewer bands, or raise the bottom of the range. I find it valuable to repair the low frequencies too in speech (certain lip-closing thumps and rattles in the sibilants get removed) but the lower frequency bands do take more computation time. De-esser performance is not too bad with many high frequency bands: but very few bands may give poor results, affecting vowels.
Both tools specify a step size for precision in identifying the intervals to repair. You might trade precision for speed.
Both tools specify another time period for cross-fading of repairs around each clicky interval. Make it too short and you introduce undesirable artifacts. This setting should not affect computation time significantly.
The De-Esser has a dB threshold which is the amplitude within each frequency band, to which, the band is trimmed back. -20dB is my default setting, and -20dB and above is also what appears white in spectrogram views with default settings.
De-Clicker has a threshold in dB which is not absolute, but relative: a band must rise so much over a short interval for a click to be detected. I repeat, subtle clicks may be treated better not with a lower threshold, but with more numerous and narrow bands.
De-Clicker specifies a minimum separation between clicks. Make it too small, and more clicks will be identified, but there may be too much damage to voice when it drops in pitch. There is also a dB threshold (for the entire signal) below which closer clicks may be detected. This may help clean up crackles in de-voiced intervals and breaths. Use low-frequency rolloff effects first for better results!
De-Clicker can specify a number of times to repeat treatment, and a maximum click length in steps, which need no explanation.
De-Clicker can also just make labels around clicks without repairing them. This information may be useful, for instance, if you want to find dB threshold setting just low enough to detect certain of the clicks. DeEsser.ny (55.6 KB) DeClicker.ny (70.5 KB)
I actually ran it on the test piece I uploaded for you all yesterday and unless Iām hearing things, much to my astonishment, appreciation and gratitude, it got rid of them!
I really needed and wanted to get some recording on the e-book done tonight so I ran through a new foreword and five chapters before my throat felt like sandpaper. I shall be processing them tomorrow (today) so that will be the real test and I shall let you know.
Thank you. I gather you use the defaults. If you think you find better settings (for bands and frequency range, perhaps, reducing compute time without reducing quality of results) let me know.
I came across this forum page and the information about your two plug-ins over the weekend.
I ran them through a couple of recent recordings that I have produced and they seem to me to have made a noticeable improvement - particularly in dealing with the number of āclicksā that inevitably scatter themselves through my recordings!
I posted a link to and details of these plug-ins on the Librivox.org forum to see if anyone else would find them useful.
(Librivox is a site where volunteers record books and other publications that are in the Public Domain, which are then made freely available for downloading).
One question though is - what is the best way to give you feedback on these plug-ins?
If youāre up for receiving feedback, I could post the details on the Librivox forum if you want.
Anyway, thanks for putting in the time and effort to produce these plug-ins. I think that they could be really useful tools to have at hand for cleaning up recordings.
I have also very recently started the Audacity Voiceover Usersā Group on Facebook, specifically to discuss the technicalities of how to get the best results out of Audacity for spoken word. I donāt pretend to have all the answersā¦ which is why I made the group so I can learn! Most of the members have come from the Audiobook Creation Exchange, but I would like to attract LibriVox people to it too.
Steve, I am trying to join the librivox forum. They seem vigilant about spammers so Iām not in yet. Havenāt heard back from your gatekeeperā¦ update, just got in!
There seem to be suggestions that the De-Esser is very good. Is the De-Esser also good for music vocals? Does it need more development? Would you want it to be published on the Audacity Wiki if so when?
I am not sure Iām doing a smartest things mathematically with this de-esser, but it seemed to drop out of the more elaborate de-clicker as a sort of simplified special case. As for the de-clicker, I am also not sure I am doing the smartest possible things there. It all runs at a snailās pace. This is a free tool I offered to the patient.
Trebor speaks highly of your De-esser - he said it was better than Spitfish.
Not yet - Iāve no use for a de-esser most of the time. But tools donāt have to be mathematically perfect to be on the Wiki. And if they are on Wiki there is more chance of getting feedback from multiple use cases to improve them.
Well I donāt really know what ārealā de-essers do, but mine could be described as a multi-band brickwall limiter, with many narrow bands in the default settings, which account for the slowness but perhaps also for precision of results if you can wait for it.
I hear it said that āsibilanceā is eqād down at 7 or 8 kHz, but I find that s sounds in my speech recordings sometimes have unpleasant piercing whistles in them that are much lower, between 3 and 4 kHz. You can easily see such things in spectrogram. I wrote something that does the program equivalent of scanning the spectrogram for such bright bits, and then fixing each one precisely over the right interval with the right gain using the Nyquist eq-band function.
No Joke : your de-esser is the best Iāve encountered, it allows far more precise control of sibilance than any other Iāve seen ā¦
SpitFish canāt do that.
If itās use is extended to processing stereo music then it will need the option of linking the stereo channels , rather than processing each channel independently , ( if they are not linked then there is an unwanted stereo flutter effect ).
Your suggestion for stereo is a good one one. Have you observed that defect? I havenāt, I work only with narration, all mono, no mixing. Will you test improvements if I write them? Or send me a test case?
Attached is my own pet example of a bit I captured from a professionally produced audiobook that I enjoyed very much for story, but the sibilance was a knife in my ears sometimes. This especially painful s whistled at about 3400 as you see, not the 6kHz and above range. This was my acid test case as I developed. The āafterā with default de-esser settings is right.
Your picture shows āmulti-band brickwall limiting with many narrow bandsā in action. Find just the white stuff and hammer it down. The danger is changing things too much and making the voices lisp. I have used this on enough hours of voice to believe it does not not.
Do use use default settings on the spectrogram and do you understand the colors? Every frequency band at -20 dB or higher appears white in default display settings. And -20 is exactly what I limit to in the toolās default settings.
Trebor, I see you use even more bands than I do and a higher top frequency. You must be very patient! My work has everything above 11.5 kHz stripped out by other parties so I donāt worry about that.
You probably know this: if you choose Spectrogram view, then the bands you specify do not correspond to equal divisions in screen height. But choose spectrogram log F instead ā then they do correspond to equal heights. And thatās how a graphic equalizerās sliders work too.
If there is a rule in the code that colours relate to a specific dB level at default display settings, that might be nice to put in the Manual. All we know there is that blue is least energy and red and white are most energy.
