The value entered in this field will remove from the spectral and roughness
analysis of each signal portion all spectral components with amplitudes below
this value.
For synthesized complex signals, whose unknown1 spectra contain only
discrete sinusoidal components, this parameter may be left equal to 0.
For all other signals, multiple analyses that experiment with the amplitude
threshold value will be necessary, in order to get roughness results that can
support valid roughness comparisons among the sound files analyzed (see at the
bottom of the page for tips).
The need for non-zero spectral amplitude threshold is due to the fact that, for natural signals, several of the
up to 50
spectral components returned from the analysis will represent signal and/or
analysis noise, and will have very low and similar amplitudes. Since sine-pairs of equal
or almost equal amplitudes result in maximal calculated roughness (see
the roughness model), the roughness
contribution of 'noise' components will overestimate the total calculated roughness.
As a result, for example, the roughness calculated for a
quiet, solo violin passage may be higher than the roughness calculated for a
strong dissonant
chord performed by an entire string orchestra. To avoid such invalid results,
the amplitude threshold value is designed to remove the 'noise' components
(components with very low and almost equal amplitudes) from the roughness
calculations, while retaining the components that contribute to the signal's
perceived roughness (see below).
Tips:
After determining whether or not you will be using spectral normalization, run
the first analysis of all sound files (or sound-file portions) you will be comparing, with consistent
normalization settings and 0 amplitude threshold. This will calculate roughness
values based on the 50 strongest spectral components of each sound file/portion.
Make a note of the overall highest lowest amplitude from all the analyses that
returned 50-component spectra and
re-run all the analyses using this value as the amplitude threshold.
For a more sophisticated (and complex) amplitude threshold selection, search the
spectra returned from the 0-amplitude-threshold analyses for components
with an abrupt drop (≈≥50%) in amplitude, followed by 3 or more components with
similar amplitudes (within ±≈5% of one another). If such components exist, re-run the relevant analysis,
using a value just above their amplitude as the amplitude threshold.
We will be
working on an algorithm automating amplitude threshold selection for the next
release of SRA.