Gianni Pavan wrote:
>
> At 13.07 21/06/2002 -0400, you wrote:
> >When using sonograms be sure and be aware that the math involved in
> >producing sonograms produces artifacts in the display. In particular a
> >lot of what looks like harmonics off the louder parts of the call is not
> >there.
> >
> >Walt
> >
>
> Walter,
> if harmonics appear on a spectrogram I think it is because
> something is in the signal, maybe due to distorsions (overloads) in the
> input stages of many small DATs (the "brickwalling" effect of the small
> SONY DATs is well known...) or in a poor sound card. I can't believe
> harmonics are generated by the math involved in computing digital
> spectrograms. Of course FFT computation should be done with the right
> algorithm and with large dynamic (I mean using float math instead of
> integer math and arranging scaling to avoid overflows). Although, it is
> true that the time-frequency uncertainty principle may cause confusion when
> analyzing pulse trains: they may appear as continuous tones with lots
> of harmonics when the pulse interval is shorter than the analysis window.
> In such a case waveform inspection and a correct tuning of analysis
> parameters (fft size, window size, window shape, scanning step, dynamic
> range of the display) may help in revealing or at least in clarifying a bit
> more the real structure of a signal.
I'm certainly aware of lots of ways harmonics creep into recordings
through failings of the equipment. And those are a whole different subject.
Nor am I contending that all harmonics in sonogram displays are a result
of the computations. After all, some are real.
The trouble is in real life most of us are using programs that don't
have all that much adjustment. And even if we did we may not be
mathematically savvy enough to make intelligent choices. And the other
nasty part seems to be as you adjust parameters, at the same time you
don't get this strange stuff your display also loses it's resolution and
clarity.
In my current "best" sonogram producing program, Spark XL, I have four
choices for FFT size from 512 through 4096 and usually choose 4096. I
also have several steps in scrolling rate (timescale) and have almost
infinite control on the frequency range (in linear or log scale) and the
correspondence of color to dB level. And there is a quality adjustment,
for which I'm not sure what math is involved, but set it better quality
and the sonogram will show a little more detail. The sonogram window is
of fixed size. I can, of course, filter the sound, or resample it. This
program shows such "harmonics" off the most intense part of the call for
some calls when otherwise adjusted to optimally create a sonogram.
In any case there's a quick way to see which it is, real or math. Pass
the file through a good cut filter to slice off all the frequencies from
just above the primary. Or lacking such a filter resample the file at a
rate that will cut off at the right point, then resample back up to the
original rate and do the sonogram. If that does not change the
"harmonics" showing up on the display, then you can be fairly sure it's
the math and not something introduced earlier in the recording or
transfer process. And by math, I don't mean the formulas we think we are
calculating but what the computer is really doing. Add, subtract and
fakes multiply and divide, computers don't really do exotic math.
I've seen this sort of thing with every sonogram program I've used. Long
ago when I first started working with computers writing computer models
I learned that what we thought went on and what does are not always the
same with computers. Which is probably why the word algorithm is stuck
in there when talking about what a computer does. The sonograms seem to
be no different in this regard. Still, computer generated sonograms beat
the old mechanically generated ones by a wide margin. And provide a
wealth of information.
One last thing, what I'm talking about is most likely if the highest
intensity parts are close to the 0dB clipping point. (note I'm talking
about sounds that were not clipped) Like within 10-15 dB. So they can be
minimized by keeping the sound from that area. Readjusting the gain of a
digital file can often stop them.
> Just to complete my thought, I should say that the sampling rate often is
> not high enough to reveal the high frequency structure of a sound: a pulse
> train with 1/10000s pulse interval appears like a perfect sinusoidal tone
> at 10 kHz if sampled at 32kHz (with good anti-aliasing filters, of course);
> a 48 kHz sampling reveals a 2nd harmonic... only a very high sampling rate
> reveals its "pulsed" and not continuous nature.
Beyond what most nature recordists have for equipment. Off into the
realm of Bioacoustics gear. And to truly find that it's pulsed, all the
input gear has to be up to it. My thought is that probably the gear used
by nearly all members of this group will feed a sinusoidal signal into
the recorder from the pulse.
But the thought is something to pay attention to. A frequency of half
the sampling rate is mangled as to waveform and thus sound quality. If
your frequencies of interest fall into the top third of the frequencies
recorded with digital a lot of the time, consider going to a higher
sampling rate. That is not justification for the current fad to jump to
higher sampling rates than the 44k CD rate. The upper third of the
resulting 22k range is not occupied by very much of what we are trying
to listen to, which is lucky because our ears are also very poor at
picking up sound in this range.
Walt
________________________________________________________________________
________________________________________________________________________
|