Continuous Noise Filter (Part 1)

[Craig Maier]

Continuous Noise Filter (Part 1)


The Continuous Noise Filter is useful for reducing background "Hiss" and other constant noises from a recording or from a noisy AM or FM radio transmission. Magnetic tape recordings and phonograph records all contain a residual background “Hiss” type of noise that can be reduced with this filter. It is referred to as a "Continuous" noise filter (or CNF) because unlike impulse noise, hiss is present at all times. When adjusted properly, this filter can almost completely eliminate all residual continuous noise from a recording. However, it is easy to overuse this filter and leave the recording sounding dead and lifeless, and also introduce digital artifacts into the signal (music or speech).

The Continuous Noise Filter has 4 modes of operation to choose from:

• [*=2]Normal CNF Mode [*=2]Spectral Subtraction Mode [*=2]Auto Spectrum CNF Mode [*=2]Forensics AFDF Mode

To use this filter in Normal CNF or Spectral Subtraction modes, you must first take a sample of a section of noise from your file. This noise template will then be used by the algorithm to discriminate between what constitutes noise and what constitutes music or speech during the filtering process. It is important to sample a section of the .wav file that does not contain any music/speech so that the filter does not remove signals that contain musical or speech information. Of course, it can be used in the opposite sense too. If you sample on a sound that you do not want other than noise, it will attenuate that sound from the final result.

Note: We strongly recommend trying the Artifact Suppression mode when using this filter; its performance is often much better than standard mode and allows for a more aggressive use of the Attenuation Control without the production of digital artifacts.


The Continuous Noise Filter
The CNF graphically shows a frequency spectrum of the sampled noise in red (sometimes referred to as a noise-print or noise fingerprint). This spectrum represents the amount of noise at each frequency band in the recording. The blue line represents the filter design that has been created by the software program. You can use the mouse to move the blue threshold line to tailor the kind of noise reduction that the filter performs to your taste.

This filter should only be used on recordings that have little or no impulse noise, or on recordings that have already been processed through the Impulse Noise Filter(s) in order to minimize the possibility of producing digital artifacts. When operating this filter in the Auto Spectrum CNF Mode, the system will automatically find and modify its own noise fingerprint on-the-fly. Therefore, there is no need to manually take a noise fingerprint when operating in the Auto Spectrum CNF Mode.

This is one of two different types of non-linear filters that can be used to reduce noise from a signal source. Like the Dynamic Noise Filter, it is useful for reducing "Hiss" from a recording or from a noisy FM radio transmission. However, unlike the Dynamic Noise Filter, it will also reduce lower frequency noise. When adjusted carefully, it can almost completely eliminate all residual noise from a recording. However, when compared to the Dynamic Noise Filter, this filter is a bit trickier to adjust so as to avoid the introduction of digital noise artifacts into the Destination .wav file. It also can have some detrimental effects on the "presence" and the musical transient content of a .wav file when not properly adjusted.

This filter takes a sampling of your file and converts it into the frequency domain utilizing a Fast Fourier transform. Next it marches along to the next time interval and performs another Fourier transform. It keeps repeating this process until the entire .wav file is converted into samples which are no longer representing the time domain, but strictly represented in the frequency domain, with the appropriate Voltage, phase and frequency co-efficients for each window contained in memory. The entire audio spectrum is divided into up to 8,192 bands by a 16,384 point fast Fourier transform (FFT) algorithm (the resolution of which is user selectable). When a signal in a particular band exceeds a threshold (when operating in the Normal CNF or the Auto Spectrum CNF Modes) that you can define graphically, then that particular band is allowed to pass its signal from the input of the algorithm to the output of the algorithm. Lastly, the entire file is then re-converted back into the time domain via an inverse fast Fourier Transform. So effectively the only time during which bandwidth is provided at any of the selected number of frequency buckets is when there is a useful signal present in that portion of the audio spectrum. Otherwise, the remaining frequency bands are attenuated to a varying degree (depending on the Attenuation Control setting).

Insert Additional Control Points onto the Continuous Noise Filter

By simply right clicking anywhere on the Frequency Spectrum of the Continuous Noise Filter, you can insert your own additional Control Point(s) onto the graphical display. This can help increase the flexibility and power of the filter. Upon right clicking of your mouse, the menu that appears allows you to add a point, delete a point, or reset the control point count to the factory default setting of 11.

The following is a summary of the control parameters functionality and range of adjustment provided by the Continuous Noise Filter:

• Attack Time

This is the time required for any of the filters to "open up" on the leading edge of a signal that exceeds the threshold line on the spectral graph.

This represents the time constant normalized value at 1 kHz. The time constant for filter frequencies operating above 1 kHz will be shorter than the setting, and the time constant for filter frequencies operating below 1 kHz will be longer.

(The Attack time constant value is weighted with a -1 slope across the audio spectrum.) Small values of attack provide excellent transient response, while long values provide a minimization of digital artifacts produced by the system. A good value to start with for the Attack Time parameter is around 100 Milliseconds. The total range of adjustment for Attack time is 1.0 to 300 milliseconds. Smaller settings will improve transient response but allow more digital artifacts to pass through. Larger values will decrease transient response, but will minimize the production of artifacts during the noise reduction process.

• Release Time

This is the time allowed for any of the filters to "close down" or "decay" following a signal that falls below the blue threshold line on the spectral graph. All remaining characteristics of the Release time Constant are the same in nature as the Attack time Constant. A good value to start with for the Release Time parameter is around 200 Milliseconds. If there is too much fast filter "breathing" (a digital artifact that is also sometimes referred to as “pumping”), lengthen this time until you are satisfied with the result. The total range of adjustment for Release time is 1.0 to 1000 milliseconds.

• Attenuation

This control sets the degree of attenuation for signals that are present and below the blue threshold line. The greater that one sets the Attenuation control, the greater will be the degree of noise reduction. However, the greater the degree of noise reduction achieved, the greater will be the loss of the sense of "Ambiance" on the resultant recording. So you must make a careful judgment as the correct tradeoff between noise reduction and ambiance for the material you are dealing with. A good value for the Attenuation parameter is around 10 dB to start with. If there is too much loss in signal ambiance, decrease this value. If you desire more noise reduction, increase this value. If you increase the attenuation too much, you will begin to introduce some digital aliasing artifacts into the Destination .wav file. The total range of adjustment for Attenuation is 0 to 100 dB.

• Threshold (Blue Graphical Threshold Line)

This feature controls the threshold value above which a signal at a particular frequency must exceed before it is passed through to the output of the algorithm without attenuation. It works in conjunction with the "sample noise" button. Although the continuous noise filter has up to 8,192 discrete frequency bands, it would be inconvenient to have to set each of them. DCArt10/DC Forensics10 provides up to 256 inflection points (shown as blue dots connected by blue lines on the graph of amplitude vs. frequency) that can be moved along both the frequency and the amplitude axis. The software will place these inflection points automatically at approximately 10 dB above the noise floor after you perform the "sample noise" function. Try these settings first to find out if the results are acceptable. Thereafter, if there seems to be some noise that needs more attenuation at a particular frequency, adjust the threshold upwards utilizing the left mouse button at the frequency of interest until you are satisfied with the results. The graphical threshold line is adjustable "live" when preview mode is enabled. If you want more than the 11 inflection points provided by default, double click using the left mouse button while pointing at the desired position on the graph. To remove inflection points, use the right mouse button in a similar manner.

• Threshold Control Grouping
  1. 1. Up & Down "Shift Threshold" Control - This feature allows you to globally shift the entire threshold line up or down independent of frequency. The feature consists of an up and down arrow box. It is operated via the left mouse button. The amplitude resolution of the control is 4 dB / click. After clicking on either of the arrows, you will see the entire threshold line shift in the direction of the chosen arrow.
     2. "Reset" Control - This button will restore all of the threshold line inflection points to their original default settings.

• Keep Residue Function

When enabled, the "Keep Residue" function will allow you to "preview" (hear) or process to the Destination Workspace the algebraic difference between the Source File and the Filters Output. In essence, you will be listening to the noise which would have been removed from the Source File had this function not been enabled. It is sometimes useful (via this function) to be able to hear just how much of the real audio signal along with the noise components that you are removing from the source signal. However, it is only fair to warn the user not to make final adjustments using this feature, as that technique can be quite deceptive. It is always best to optimize your parametric settings for the best results while listening to the actual processed filter output signal (i.e. "Keep Residue" function off).

• Artifact Suppression Mode

This mode of operation reduces the level of digital artifacts (sometimes referred to as “the birdies”) produced by the CNF during its noise reduction process, especially when it is being used aggressively. It can also be used to effectively reduce some forms of inter-modulation distortion (IM Distortion) from a recording (like that “raspy” sound found on some distorted 45 RPM records - - - especially on vocals). When operating the CNF in Artifact Suppression Mode, the Attack and Release functions are eliminated in that the routine operates independently of those parameters. The “Attack” control will become grayed out and the Release control will revert to a new mode of operation simply called “Artifacts”. The artifact suppression mode not available when operating in the Auto Spectrum CNF mode. In artifact suppression mode, use the CNF as you normally would, including the taking of a noise print sample in the other two modes of CNF operation (normal mode and spectral subtraction mode). The Artifact Suppression control affects the degree to which the system attenuates digital artifacts with higher settings providing a more aggressive action. You should note that higher levels of “Attenuation” are achievable in Artifact Suppression Mode compared to non-Artifact Suppression mode. 40 to 50 is a good setting to start with for the Attenuation control with an FFT size set to 4096. Experiment with FFT sizes on each side of the recommended value to find the optimal result on any particular file. 200 is a good setting to start with for the Artifact Suppression Control. Adjust this control upwards for an increased artifact reduction effect and downwards for a reduction in the removal of audio material (ambience and transients, etc). Use this control to find the best balance between those two sonic characteristics of the particular material that you are working with. Note that the optimal settings for the mentioned controls are substantially source material dependent. Using the “Keep Residue” mode will allow you to monitor how much audio material is being removed from the signal by the CNF Artifact Suppression system.

The Artifact Suppression system is not functional in Auto Spectrum mode but does operate in Normal mode, Spectral Subtraction mode and Forensics AFDF mode. Also, it is to be noted that fairly high FFT sizes produce more optimal results with this system. FFT sizes below 1024 are not recommended with the Artifact Suppressor because the system is less effective due to the poor frequency resolution associated with reduced FFT count. It is worth noting that the Artifact Suppression mode requires much higher levels of CPU resources and thus takes around 4 times longer to process a given file compared to the normal (non-Artifact Suppression) mode.

• Smoothing Mode

An alternative to Artifact Suppression mode is Smoothing Mode. The smoothing checkbox applies some additional signal processing that reduces the digital artifacts or “Musical Noise” that may be heard when applying the Continuous Noise Filter by averaging adjacent frequency bins within an FFT. This feature is most useful on very noisy recordings with high levels of surface noise such as old 78s or extremely noisy Forensics situations. When using the smoothing function, you should be able to increase the level of noise reduction without introducing digital artifacts by about 3 dB or more. The tradeoff (and there are always tradeoffs), is that the frequency selectivity is reduced in this mode (sometimes resulting in reduced bass response). An example of where the smoothing function would not be useful would be when you are trying to remove pure tones or steady state buzzing sounds where you need to maintain a high degree of frequency selectivity. Please note that this function can’t be used in conjunction with the Artifact Suppression mode. It is an alternative to Artifact Suppression mode.

• FFT Size (Resolution)

Choose between 32, 64, 128, 256, 512, 1,024, 2,048, 4,096, 8,192, and 16,384.

The frequency resolution of the Continuous Noise Filter can be adjusted. This parameter determines the number of frequency bins used by the FFT algorithm. The actual number of frequency bins produced is the FFT Size divided by two (since the FFT produces data for both the real and imaginary planes). Large values of resolution produce the largest degree frequency selectivity and thus high degrees of noise reduction, while the best time domain transient response will be realized with smaller FFT values. Put another way, there is a tradeoff between frequency resolution and time resolution and they are inversely related to one another. You will have to experiment with the various binary weighted values to determine the best resolution for the material that you are dealing with. Listen for the best levels of noise reduction attainable while minimizing any digital artifacts and yet maintaining good musical transient response on things like rim shots on drums, and other percussive instruments.

Important Note:
This parameter cannot be adjusted while the filter is previewing or running.