Continuous Noise Filter (Part 2)

[Craig Maier]

Continuous Noise Filter (Part 2)

• Overlap (Window Overlap – Forensics Version Only)

This allows you to choose the percentage of overlap between FFT windows. The range for this adjustment is 30% to 50%. Often, reduced values of % overlap will yield better results in terms of digital artifacts produced, but will take longer to process. Conversely, larger values of overlap will produce more digital artifacts, but take less time to process. Start with 50% and lower it order to achieve faster sonic transients tracking. Note that this feature is only available in the DC Forensics10 Audio Laboratory version of the software. The overlap parameter is fixed at 50% in the DCArt10 version of the software and is not visible to the user.

• Spectral Subtraction Mode In The Continuous Noise Filter

This changes the mode of the Continuous Noise Filter into a spectral subtraction system making it especially useful for some Forensics noise reduction applications. As with the normal mode of the filter, you will still need to take a fingerprint of the noise when operating in Spectral Subtraction mode. The amplitude signal of the resultant FFT is then subtracted from the entire .wav file signal thereby providing rejection of the highlighted fingerprinted signal. The offending signal could be an air conditioner, city noise, automobile noise, etc. To adjust the depth of effect, use the attenuation control. The attenuation slider sets the maximum amount of attenuation as an absolute number (as in all of the CNF modes) and in Spectral Subtraction mode the threshold shift scales the overall gain of the noise sample. Threshold and attenuation are not equivalent parameters because of the fact that the maximum attenuation value is absolute but it is not a maximum delta value between the signal and the sampled reference.

Important Note:
If the parameters for the Continuous Noise filter are set incorrectly, it has the propensity to produce extremely strange sounds (digital artifacts) - - - some of them quite comical in nature. If you hear these "birdies", back down on the attenuation setting and / or some of the graph inflection points and they should disappear. If not, increase the Attack and Release time values.

• Auto Spectrum CNF Mode

Sometimes, it may be desirable to have the Continuous Noise reduction system determine its own noise fingerprint “on the fly.” There are three possible reasons why you might be interested in using this mode:

• [*=1]It simplifies the operation of the Continuous Noise Filter [*=1]There may be no discernable quiet passage on the source material from which to “Sample the Noise Fingerprint”. [*=1]The Noise Fingerprint may be changing in its noise distribution dramatically throughout the .wav file.

The Auto Spectrum CNF Mode solves these problems by calculating the noise fingerprint “on the fly” and updating it on a continuous basis. It is capable of performing this function even if you can’t see or hear a totally quiet passage in the target .wav file. The mathematical routine utilized to create this fingerprint is more sensitive than your ears and can always find a noise fingerprint. This system may not be quite as effective for removing noise compared to the other two modes of the CNF operation, but its convenience coupled with its ability to adapt to changing noise environments sometimes outweigh its noise reduction performance limitation. To operate the CNF in this mode, merely preview it and adjust (primarily) the attenuation, attack and release times for the best sounding results. As with all of the Diamond Cut filters, the “Keep Residue” mode can be useful as an adjustment tool. In “Keep Residue” mode, tune for maximum noise and minimum signal while operating in “Preview” mode. The Auto Spectrum CNF Mode has the following controls and displays:

• [*=1]Attenuation (start with a setting of 25) [*=1]Attack Time (start with a setting of 20 mSec) [*=1]Release Time (start with a setting of 50 mSec) [*=1]FFT Size (start with a setting of 8,192) [*=1]Smoothing Checkbox (on) [*=1]Overlap % (50 %) [*=1]Update Spectral Display (on/off selector checkbox - - - your choice)

The Auto Spectrum CNF Screen

Forensics Adaptive Frequency Domain Filter (AFDF) Mode
(Forensics Version Only)

The Forensics Adaptive Frequency Domain Filter (AFDF) is a variation on the basic theme of the Auto Spectrum CNF filter mode. The primary difference is that the Forensics AFDF is optimized for Forensics oriented files and not “High Fidelity” files. It is found as one of the modes under the CNF filter or directly via an item called “Adaptive FD Filter” found under the Forensics Menu*. It has a faster response time compared to the Auto Level function and also a narrower effective bandwidth while producing higher levels of noise reduction at the expense of potentially producing higher levels of digital artifacts. The AFDF Filter is “Adaptive”, which means that it will automatically adjust itself to varying noise environments. This filter has a Time Domain twin sister found in the Forensic Menu, which is simply called the “Adaptive Filter”. The AFDF can produce more attenuation of noise compared to the Time Domain Adaptive Filter but can also produce more spurious digital artifacts. The Time Domain Adaptive filter does not produce spurious digital artifacts, but can produce an echo delay sound as an artifact. We provide both types of Adaptive filters since they exhibit somewhat differing sonic characteristics. Experimentation is the best way to determine which filter is the most effective on a given Forensics audio file. Also, a combination of the two used together in cascade in the Multi-Filter has been shown to work some magic that otherwise would not be possible on some extremely noisy forensics files. In general operation, the controls and the display graph for the Forensics AFDF is much the same as that of the Auto Spectrum CNF. Refer to it for details.

*Note 1: When you need to use the AFDF in the Multifilter, you must access it via the appropriate CNF mode inside in the CNF.

Note 2: The “Green Zone” spans for the various slider controls will vary depending on the chosen CNF mode and/or whether Artifact Suppression is active or inactive.
Continuous Noise Filter Procedure (Tutorial for Normal CNF Mode)

The Continuous Noise Filter is one of the most mathematically complex of all of the DCArt10/DC Forensics10 algorithms. It will, therefore, take the longest amount of processing time to complete its calculations. This algorithm will benefit the most from the use of a high clock rate computer. This filter is also the most difficult filter to use correctly. Aliasing artifacts can be produced when the settings are not correct for the particular .wav file you are attempting to "de-noise." The first time you use it, it will be worthwhile to spend about an hour playing around with it in order to become familiar with its behavior.

1. Highlight a quiet portion of the Source .wav file. Often, this sector will be found at the beginning or at the end of the file, as with the lead-in or the lead-out groove sector of a record or the lead-in portion of a tape recording. The idea here is to capture a section of noise only, but no signal. This will become the noise floor baseline for the subsequent operation of the Continuous Noise Filter. Only about one or two seconds of noise is needed.
2. With the left mouse button, click on the "Filter" menu.
3. Next, click on "Continuous Noise”.
4. When the Continuous Noise Dialog Box appears, click on "Sample Noise."
5. Some calculations will be made in the ensuing moments. When they are complete, a graph will appear showing the Amplitude (in dB) versus the Frequency of the .wav file noise floor. This graph represents the Noise Print of the file on which you are working.
6. The measured sample noise spectrum is shown in red. The noise threshold value versus frequency is shown in blue. You can set the blue graph threshold value, although DCArt10/DC Forensics10 will automatically choose some settings for the threshold limit line that is a good place to start.
7. If you choose to change the graphical threshold contour, follow the procedure outlined in steps 7 through 10. Using your mouse, place the pointer on the left-most blue threshold marker on the graph (one of ten blue dots).
8. Depress the left mouse button and move the dot either up or down so that it remains somewhere above the red line graph at the bottom end of the spectrum. The higher this line is from the red line, the greater will be the degree of noise reduction at frequencies near this particular dot. If the dot is placed below the red graphical line, no noise reduction will be applied to these frequencies. This is sometimes the preferable setting for the blue threshold line near the bottom end of the audio spectrum (below a few hundred Hertz).
9. Next, move the next blue threshold marker just to the right of the first one, and using the mouse, set it somewhere above that particular frequency on the spectrum graph.
10. Repeat this process until all 11 threshold markers are located somewhere above the "noise floor" graphical representation of your .wav file. Now the blue line should be located above the red line at all frequency locations. Note that the best contour can only be achieved by not only moving the markers along the vertical axis, but along the horizontal (frequency) axis as well.
11. Set the "Attack" time initially to 20 milliseconds.
12. Set the "Release" time initially to 50 or 100 milliseconds. (The "Release" time constant should always be set longer than the "Attack" time constant for a realistic sounding operation of the filter.)
13. Set the "Attenuation" control initially to 10 dB. (Higher numbers results in higher levels of noise reduction.) Too much noise reduction will produce digital artifacts and detract from the "ambiance" of the recording.
14. Highlight the portion of your .wav file on which you desire to apply the Continuous Noise Filter. (You may choose to highlight the entire file or any portion thereof.)
15. Preview (or Run) the Filter.
16. Listen to the result and then determine which parameters need modification. If there is a "lagging" response to audio signals, decrease the "Attack" time. If there is a "swell" of noise following an audio crescendo, decrease the "Release" time. If a portion of the audio spectrum is sounding dull, lower the "threshold" line at the frequency of interest. If a portion of the audio spectrum is sounding noisier compared to the rest, then raise the "threshold" at the frequency range of interest.
17. When you are satisfied with the results, Run the .wav file (in its entirety) to achieve the final processed results in the "Destination" workspace.

Note 1: The threshold line inflection points can be adjusted "live" when you are running the filter in "Preview Mode." You will be able to hear the effects of modification that you make to the threshold line almost immediately.

Note 2: Sometimes, record albums that are in excellent condition exhibit only one type of noise which is called “Rumble”. Rumble comes from the record mastering process as well as from the bearings in your turntable. Generally, one uses a High Pass filter to remediate this problem, but that approach also cuts into the low pass portion of the audio signal. There is a special preset in the CNF called “Dynamic Rumble (Only) Filter” which will remove this type of noise without damaging the deep bass on the recording or affecting any other signal above 90 Hz. This filter is more effective on Rumble noise than the standard High Pass Rumble filters found elsewhere in this software program.

Note 3: The most common issue with the use of the CNF is setting it too aggressively. It’s nice to reduce noise, but not to aim to completely eliminate it, as that approach will produce digital artifacts.