Adjusting DeHiss/DeNoise Parameters

There are five parameters you can use to control the process of sound renovation - one primary parameter, indicating the rate of noise reduction (gain in DeHiss Plus and weight in DeNoise Plus) and three secondary ones which can be used to adjust degree of spectral smoothing (smoothing range), degree of high frequency attenuation (frequency carving) and the size of the local analysis frames (frame size). We recommend you start tuning DeHiss Plus or DeNoise Plus by adjusting only the primary parameter and keeping the other processing parameters fixed on their nominal values. As long as it becomes clear what are the most appropriate gain/weight values, you can further try to improve the results by fiddling with the remaining tuning knobs.

Gain

(for DeHiss Plus only)

The gain coefficient, deciding upon the amount of noise the program attempts to reject from an audio file, is the most important "knob" you have to tune when using DeHiss Plus (you can pick one of the five standard gain values from the gain list; for fine tuning use the slider). The best results are obtained, of course, if the assumed noise intensity matches the true one. If you set the noise intensity at a too low level the renovated signal will contain an audible residual noise; if the level is too high the sound may become dull, muted and/or distorted. Be aware that since the gain is expressed in logarithmic units, even small changes in its value may cause significant changes in the renovation results.

Weight

(for DeNoise Plus only)

Weight in DeNoise Plus plays a similar role as gain in DeHiss Plus. If the noiseprint is taken from the processed audio file, try the values from the normal range. Otherwise, you may have to consider larger (up to 20) or smaller (down to 0.01) values to compensate for the difference in noise intensities between the noiseprint and a given recording. When adjusting the weight use the same rules of thumb as those suggested for choosing the gain in DeHiss Plus.

Smoothing range

A special spectral smoothing technique is used to reduce the residual noise. If the degree of smoothing is too low the renovated signal will sound rough and/or mechanical. In extreme cases it will be corrupted by a specific disturbance called "musical noise" - a random combination of different tones (as if someone was casually striking different piano keys). If the smoothing range is too large the restored signal may sound hollow and lifeless.

Frequency carving

Lowpass filtering is perhaps the simplest way of reducing broadband noise. Even though DeHiss Plus and DeNoise Plus suppress the high frequency components of the processed sound, you will usually get improved results if some extra attenuation is forced into the range of high frequencies. The degree of high frequency attenuation should be chosen in accordance with signal characteristics, e.g., to preserve rich sound of some musical instruments (a church organ, for example) you have to use this tool with caution. On the other hand, without frequency carving the renovated signal may sound mechanical and/or the residual noise may be too emphasized.

The choice of the rate of frequency carving should depend on the sampling frequency (you can safely use large and very large rates for 44.1 kHz recordings but in most cases you should pick small or normal rates for the 22.05 kHz files) and on the noise level (for very noisy audio files, large rates are advisable).

Frame size

In the course of processing, the signal is divided into the partially overlapping analysis frames. Longer frames yield more uniform renovation results - use them when the signal to noise ratio is small. For high and moderately high signal to noise ratios try shorter frames because short frames will help you preserve freshness and liveliness from the original sound.

Use shorter frames to process speech signals and longer ones to deal with songs and instrumental music.

Dynamic noise reduction

When a constant noise suppression rate is applied to all parts of an archive recording, the results are often unsatisfactory. For example, suppression rates that are adequate for silent parts of the recording often turn out to be too small for louder parts and produce an annoying “noise aura”. And conversely – when noise reduction is tuned to loud parts of the recording, its silent parts become lifeless and muted. To achieve best results, the level of noise reduction should be made dependent on the volume and spectral content of the currently processed fragment. Both DeHiss Plus or DeNoise Plus allow you to use such adaptive noise reduction mechanism. To turn dynamic noise reduction on, check the box situated in the Dynamic noise reduction group. Then adjust two parameters - Noise floor and Adaptation rate – to obtain the best restoration results.

The Noise floor parameter sets the volume of the signal above which the adaptive mechanism is active. For example, if Noise floor is set to 500, the constant adaptation rate is applied to fragments with mean volume smaller or equal to 500. For the remaining fragments the noise reduction is increased above its nominal level – the Adaptation rate coefficient controls the amount of this extra suppression (the larger the adaptation rate, the wider the dynamic range of noise reduction).

NOTICE

When dynamic noise reduction is turned on, very good results can be usually obtained for small values of the restoration parameters (gain/weight, smoothing range, frequency carving and frame size) which has some clear benefits – see below.

General guidelines

The following two guidelines, should assist you in using the DeHiss Plus/DeNoise Plus interface in a logical way

In order to achieve greater noise reduction try to increase the noise reduction gain/weight (first choice) and/or to increase the remaining processing parameters.

In order to make the restored signal sound more natural, fresh and lively try to decrease the noise reduction gain/weight (first choice) and/or to decrease the remaining processing parameters.