Denoising Process 

Since the entire process runs in real time, you can simply switch between the three noise profiles while listening to the output signal. This helps you determine which noise profile is best suited to the audio material being processed. We discovered that a shelving-like modification of the upper end of the noise profile helps to improve the results a lot. This can be done by adjusting the parameters cut freq and cut gain.

The noise reduction process in the DeNoiser PlugIn is controlled basically by just two parameters: threshold and reduction allowing an easy search for optimal results depending on the given input signal. The remaining parameters are for the fine tuning.

The frequency spectrum of the input signal as well as the spectrum of the noise taken out during the denoising process can be followed in the noise scope. You can make an estimation of noise amount being removed based on the noise level indicator. Although for accurate examination of the frequency spectra, we recommend using the Analyzer PlugIn. The Analyzer PlugIn can display the input and output signals of the Analyzer PlugIn as well as the noise profile applied to processed audio material.

By setting the Analyzer PlugIn to DeNoiser in (red), DeNoiser out (green), and the white line to the DeNoiser noise profile, you can intuitively follow the effect of the denoising process on the processed material. The noise profile (white) marks the threshold border, above which no noise reduction is applied. The threshold parameter moves this noise profile up and down and can be used to set up the profile just above the background noise level. For a given threshold, the second parameter, called reduction, controls the amount of the spectral components removed below the chosen noise profile.

A good starting value for threshold is to set the noise profile just above the background noise level (approx. 10 dB). A subsequent increase of the reduction parameter should significantly reduce the background noise. If noticeable artifacts in the form of so-called singing or robot-like sounds appear (time aliasing phenomenon), decreasing the reduction parameter and increasing the threshold level (up to about 30 dB above the background noise) usually helps. Further reducing of artifacts can be achieved by careful setting of the third slider modify.

The overall performance of the DeNoiser PlugIn is heavily dependent on the proper adjustment of the parameter in the Quality Window which are related to the frequency-domain processing. Especially for the final mix, the rules recommended for the Quality Window settings should be carefully followed.

The noise scope window displays the frequency spectrum of the input signal as well as the noise taken out during the denoising process. This allows objective control of the DeNoiser algorithm activities. 

However, especially when using a user defined noise profiles, the Analyzer PlugIn is strongly recommended. It allows visual control of the denoising process (placement of the noise profile) and precise measurement of the frequency spectrum from 10 Hz to 20 kHz with a dynamic range of more than 130 dB. The display of the Analyzer PlugIn can simultaneously show three characteristics: DeNoiser input, DeNoiser output, and the noise profile.

If the Analyzer PlugIn is not available, the noise level meter is quite useful for rough estimation of the noise amount removed from the input signal. If the background noise varies greatly, there is no general rule for the practical use of the noise level meter.

The more noise to be removed, the more the operator has to be aware of possible artifacts arising from the noise removal system. These artifacts are greatly influenced by the Quality Window settings and the noise profile used. Although the noise level meter can help in objective comparison among different setups, the final judgment should always be made by carefully listening to the output signal with a high-quality speaker system or headphones.

The noise profile should ideally represent the frequency distribution of the noise to be removed from the noisy input signal. It is a kind of reference spectral horizon used by the denoising algorithm. The position of the noise profile relative to the input signal can be controlled with the threshold parameter. The Analyzer PlugIn provides an intuitive visual control of both the input signal spectrum and the noise profile.

Applying user-defined noise profiles taken properly from the audio material to be processed, the quality of the entire denoising process can be considerably enhanced. It is important that the noise profile should be taken from a part of audio material containing nothing but the noise signal to be removed. This can be done from the Noise Profile Manager dialog, which opens after clicking on the finger print button. The AlgoPlayerÖ allows a precise marking of the waveform part best suitable for taking the noise print. It can be also stored for the later use.

Playback of the area marked in AlgoPlayer starts after pressing record in the Noise Profile Manager window. If the "noise-only" part of a recording is very short (below 2 seconds), it is recommended you set up a loop for repetitive playback before recording a noise profile (see PlugIn Station help for details). The loop can be played back several times to obtain an averaged noise profile. As already pointed out, the loop should contain nothing but noise (or whatever has to be removed from the input signal). If a noise profile includes spectral components of the signal to be recovered, they will also be removed or at least lowered in the denoising process. Therefore, much care and sensitivity is recommended when preparing a user-specific noise profile.

In many cases excellent results can be achieved using just the standard noise prints: white and pink properly modified by the cut freq and cut gain parameters. We recommended to store the good working noise prints for the future use. 

Dependent of the intensity of the denoising process some artifact in form of "singing birds" or whistle-like tones can occur. In addition to mentioned before slider modify, they can be effectively minimized by proper setting of attack and release parameters. Normally attack time should be set up to the values below 0.1 seconds (e.g. 0.03 sec.), and the release time above 1 second (e.g. 2 sec.). Longer attack times than recommended can blur the signal details. Shorter release times than recommended cause cutting of the ambience, but at the cost of keeping more noise. In this case a improvement can be done by proper setting cut freq (usually 2 û 4 kHz) and cut gain (positive values).

The fine adjustment of the parameters: modify, attack, release, freq cut, and gain cut is a iterative process requiring some experience and good feeling for the trade-off between original signal, remaining noise, and artifacts. 

For forensic applications artifacts are usually not critical. The main goal there is to get as much valuable information as possible. For the speech recordings the most important parameter is the ineligibility. So, noise reduction used there is much more intensive comparing to the recording restoration applications. Therefore attack can generally be low (0 û 20 ms) to allow all signals above the noise print level to pass the denoising part. Release can normally be kept quite large (seconds), but lower values gives more abrupt noise reduction. This can be advantageous at extremely noisy material. Use spectral modification with cut freq setting cut gain below 0 (negative values) to preserve
harmonic contents of the processed audio material. An additional combination of two PEQs, one pre-FD and one post-FD, might help to extract important information. The pre-FD PEQ can emphasize particular spectral details before denoising process and the post-FD PEQ flatten the spectrum to the original one. This effect can easily be followed with difference function enabled.