[Vorbis-dev] Bounty for function to monitor an OGG-VORBIS stream

[Paul Martin]

On Fri, Aug 24, 2007 at 09:49:02AM +0100, Andrew Brampton wrote:
> From: "Paul Martin" <pm at nowster.org.uk>
> 
> >On Thu, Aug 23, 2007 at 07:51:44PM -0500, R.L. Horn wrote:
> >>
> >>Couldn't you just look at deviation from, say, white noise (or
> >>however "static" is defined)?  Since that's a frequency-domain
> >>problem, it seems like the DCT coefficients would provide all the
> >>necessary information.
> >
> >What happens if someone decides to play a piece of music that
> >contains the sound of a waterfall, waves crashing on a beach, or
> >something similar?
> 
> I don't know anything about audio processing, etc... But from what I understand 
> isn't white noise just random particles from space, if so wouldn't the entropy 
> of the static be quite high (ie random), where as the entropy for music, or 
> other structured sounds be low?

IIRC, Vorbis does encode noise (as in hiss) separately from tones
(fundamental, harmonics), so half the battle is won for you already.

A problem is that a radio station suffering loss of carrier may not
be received as a simple hiss (a pink-ish noise, because of the
preemphasis used with FM). It may, due to the FM capture effect and
lack of selectivity in the tuner, be partly picking up adjacent
frequencies or even a distant station on the same frequency,
especially in a city environment with a crowded VHF band.

Loss of carrier handling is best done in the radio receiver, by
muting the output (squelch).

One thing that could be exploited is that the hiss will tend to be
mono, as the stereo pilot tone won't be present. Cheaper receivers
may leak pilot tone (19kHz). With a sufficiently high quality
encoding, that might still be present in the encoded audio during
normal conditions. It's comfortably within the 22.05kHz Nyquist limit
of 44.1kHz sampling. Of course, no-one relaying broadcasts needs to
use anything better than 32kHz sampling as FM broadcast stereo has a
brickwall low-pass filter at 16kHz anyway.

Your suggestion of looking for a flattish spectrum is hindered by the
current trend in loudness processing at radio stations. These
processors (eg. the Optimod series) tend to flatten the spectrum,
compress peaks, etc. in order to make a station appear to sound
louder. They're also the enemy of perceptual encoders like Vorbis,
but that's another story.

-- 
Paul Martin <pm at nowster.org.uk>