Continuous noise is a common problem with recordings captured from a diversity of sources, for example from magnetic tape (cassettes, etc.), as well as from older gramophone records.
The truth of the matter
Trying to remove hiss, hum, rumble, and similar noise, can be likened to trying to remove the milk from a cup of tea or coffee. Often, the noise can be reduced without degrading the underlying audio to an unacceptable extent. Sometimes the result will sound exactly as if the noise were actually removed, sometimes an objectionable torrent of noise will be reduced to acceptable background noise. Listen to the samples (Capercaille, Ted Weems, Percy Grainger) to see what can be achieved in reducing hum and hiss using the software on this site.
Treatment methods
• Frequency domain based – used by the DeNoise application for hiss and other general broadband noise problems.
• Non-linear filter and wavelet based – used by the DeNoiseLF application for the treatment of hum, rumble, and other low-frequency noise.
Processing artifacts
Although the removed "noise" will inevitably be correlated somewhat with the desirable audio, unpleasant processing artifacts should be minimal with reasonably chosen settings. But do remember that there is no way to perfectly un-mix what has been mixed!
DeNoise – software hiss reduction
Hiss, and other mid- to high-frequency noise, can be reduced using the DeNoise application.
Hiss reduction – how it works
Suppose you had a 1024-band graphic equalizer, and that you could adjust all 1024 gain controls 80 times a second. Impossible? Quite possible! DeNoise operates as an intelligent graphic equalizer, dividing the incoming audio into 80 overlapping frames per second, automatically and continuously adjusting the gain of 1024* frequency bands. It uses algorithms that judge what will be perceived as noise, and what will be perceived as desirable audio, and applies a balancing act to suppress the noise without doing perceptible damage to the sound.
* 2048 bands if the sample rate exceeds 48kHz.
If it's not broken ...
Have you ever wondered why mp3 and mp4 can sound as good as it does? The amazing thing is that 90% of the original data has been thrown away! It's all done using the science of psycho-acoustics – the study of how we actually perceive sound. DeNoise uses the same scientific principles as these audio codecs. Algorithms from mathematical statistics first judge what might be fixed after which methods from psycho-acoustics decide between what should to be fixed and what is best left alone. Some of the noise will be masked by the actual sound – such noise is best not touched. Moreover, the DeNoise interface allows you see precisely what it is doing, and why, to help in choosing appropriate settings for each particular audio file.
Monitoring
DeNoise is designed to be automatic, but the results may be monitored using headphones or speakers. Sound output can be selected to play the incoming untreated audio, the outgoing repaired audio, or the difference, which is what has been discarded as audible noise. Processing can be stopped at any time, so that the effect of different settings can be auditioned.
DeNoiseLF – software low-frequency noise reduction
Hum, Rumble, and similar low-frequency noise, can be reduced using the DeNoiseLF application.
Why DeNoise is not the right tool
DeNoise is a spectrally-based method, dividing the audio into 1024 frequency bands at the rate of 80 overlapping frames per second. For CD quality sound, each band has a width of about 22Hz, too coarse for the accurate resolution of low-frequency noise. In particular hum, which is at a fixed and sharply defined frequency, cannot be treated this way. Improving the frequency resolution would require dealing with frames in the order of 1 second length, whereas the effective treatment of hiss and similar noise demands a frame rate in the order of 80 per second.
How DeNoiseLF works
DeNoiseLF is not a spectrally-based method. It makes extensive use of wavelets to extract and treat the low-frequency audio content in a manner which achieves perfect reconstruction of the untreated higher-frequency content. The noisy audio signal is assumed to consist of the desirable audio and the undesirable low-frequency noise, simply mixed. Advanced statistical algorithms are used to estimate the two streams and remove the noise. The DeNoiseLF interface permits you to view both wave-form and spectral information about the audio and the treatment.
Broad-band noise versus Hum
DeNoiseLF has two mutually exclusive modes of operation.
• A "General" mode suited to the reduction of noise which is widely dispersed across the low-frequency spectrum.
• A "Hum" mode for the reduction of noise at a sharply defined frequency – however DeNoiseLF is not a notch filter.