I’ve noticed a new type of noise artefact that occurs in the new “Noise Reduction” effect (Audacity 2.1 alpha) that did not occur in “Noise Removal” in previous Audacity versions.
The attached audio file has the old Noise Removal on the first version, and Noise Reduction in the second version.
In the first version there is a noticeable breath sound after the word “regularly”. In the second version, that breath sound has taken on a distinctly “harmonic” timbre that I don’t seem to be able to avoid.
Another example. After the word “meal”, the mouth and breath noises sound a little louder, but more natural in the first (old) effect and more “metallic / artificial” in the second (new) effect.
In each case I have aimed for “good” settings.
Yes the effect is subtle, but I find that the “new artefact” occurs with any recordings in which the background noise is gentle hiss and there are breaths / mouth noises that are slightly higher level than the background hiss. Any voice recordings that exhibit those characteristics are likely to produce such artefacts.
That’s to be expected. Another case would be brushes on cymbals that were at or near the background hiss level.
Comparing the spectra of the two noise-reduced samples it’s easy to see where the new NR is gating bands that the old NR has missed. We know why the old NR is missing these bands: it’s because of the the big probem that Paul identified.
I’d like to run those originals through DeNoise, which incorporates psycho-acoustic masking principles to minimize the number of bands it needs to process.
Here’s a comparison of the spectra of the breath sound in your first sample after old and new NR as applied.
– Bill
If Bill’s guess is correct that this is just a downside of fixing the “big problem” of spectral leakage, then that means, if you build exposing the advanced controls, and use “None, Hann” and two steps per window, then new noise reduction should not be worse than old for breaths.
It seems to me that DeNoise is using similar principles to that used in MP3 encoding to decide which bands of “noise only” will be masked by adjacent bands of “signal” and choosing to not treat those nearby bands. I deduce this from the documentation, and experiments with tone+noise with varying S/N. That approach might be “frequency smoothing done right”.
