"Jim Lesurf" wrote in message
...
In article , Serge Auckland
wrote:


Note that MP3 (and MP2) use psychoacoustic masking to reduce the bit
rate. This means that the audio band is broken down into many (don't
know exactly how many) individual frequency bands, with the idea that
quiet sounds in bands adjacent to loud sounds will be masked and
therefore not codec. The many filters means that a sweep will be
constantly crossing filter boundaries dynamically, and will pretty much
confuse the MPEG coder.

Why would that confuse the coder if the level of compression wasn't
excessive? I assume the sweep rate is slow enough that it will take many
'chunks' to cover the entire range.

There *was* pretty severe compression, the output from the codec was MP2 at
128kbps, so around 9-10:1 Sweep rates were 1.5, 5 or 20 seconds for a
20-20k sweep. (You may know the Lindos LA100 test set)

Don't know the correct values off-hand, but IIRC MPEG type systems tend to
use 512/1024 sample chunks and hence chunk lengths of the order of a few
tens of milliseconds. This in turn implies bandwidths per spectral
component of the order of a few tens of Hz.

That's what I remember, my MPEG theory is getting a bit rusty. Something
else to read up about!

When there was the 199? Football World Cup in France, I sold France
Telecom a large number of automatic audio test meters so FT could test
the 400+ ISDN codecs they bought to provide audio feeds to the world's
broadcasters. These ISDN codecs were being used on MP2 at 128kbps
(pretty much as DAB is in the UK). Using spot frequencies, all tests
were fine, as was audio quality (mostly speech). but as soon as the
automated sweep went through, the frequency response had so many
anomalies as to be meaningless. FT solemly carried out the tests,
failed every codec, but used them anyway....aren't the French
wonderful?........

How were the time periods of the FT measurements related to those of the
sampled chunks? I am sure you are correct in that the response will be
affected[1], but I am curious as to how large the effects might have been,
provided the level of data reduction wasn't too servere....

The frequency response dips were 10dB+. There were several dips in the
response throughout the sweep. If I did a frequency response at spot
frequencies, if I recall, 1 sec at each frequency, the response measured
flat, but if the frequencies were gliding tones, I got these anomalies.
Codecs were AETA.


[1] Partincularly as reduction would be aimed at being nominally innocuous
on a *psychoacoustic* basis, not for a spectral analysis. :-)

Slainte,

Jim

S.