Re: [Ekiga-list] Bad sound quality because of missing low pass filter

[Damien Sandras <dsandras seconix com>]

Le lundi 01 mars 2010 à 22:47 -0500, Christian Jaeger a écrit :
> Hello
>
> I'm using Ekiga and Twinkle with the Diamondcard.us service as my main
> phone service. People are telling me every now and then that the sound
> quality is pretty poor, and in fact when I listen to my voice on the
> echo service it's always been quite much worse than when using a
> "normal" phone I think. It's regardless of the codec used, all that
> are offered by Diamondcard are using 8000 Hz sampling rate. And I'm
> using a headset and have checked sound quality locally, which is just
> fine, as I can verify for example with (the bfr command is to delay
> playback, so you can hear the echo better):
>
> arecord -t raw -f dat -r 22050 -c 1 | bfr -m 33075 | aplay -t raw -f
> dat -r 22050 -c 1 -
>
> Now 8000 Hz is of course lower, and in fact when using that rate:
>
> arecord -t raw -f dat -r 8000 -c 1 | bfr -m 12000 | aplay -t raw -f
> dat -r 8000 -c 1 -
>
> it sounds exactly as in the echo test. Now one thing I know is that if
> you sample a signal, you need a low pass before the quantization, at
> half the sample rate or lower, to remove frequencies that are too high
> to be represented in the digitalized data. If the low pass is missing,
> the high frequencies are turned to noise instead. So what happens when
> a program samples data at a low frequency, will it somehow make the
> sound card tune a low pass to the correct cut-off frequency when the
> sample rate is changed? Sounds rather unlikely, and in fact the echo
> test at 8000 Hz sounds exactly like suffering from that problem: you
> can't distinguish "s" from "f" anymore, which may just be unavoidable
> with no frequency spectrum above 4000 Hz, but worse, they sound very
> loud and noisy, and for example an combined s and vowel, like they
> often appear in french (e.g. "bourgeoisie" I think, I'm not a native
> french speaker), sound really awful. Now there's a way to remedie
> this: low-pass filtering can also be done computationally, so we can
> digitize at say DAT quality (48000 Hz) and resample down to 8000
> samples/second programmatically, which is what the following command
> line is doing:
>
> arecord -t raw -f dat -c 1 | bfr -m 36000 | sox -t raw -r 48000 -s -2
> -c 1 -  -t raw -r 8000 - resample -ql |bfr -m 6000 | aplay -t raw -f
> dat -r 8000 -c 1 -
>
> It sounds more hollow/muffled now, but much cleaner. No loud noises
> anymore for sharp consonants.
>
> Would doing this in VoIP programs improve perceived sound quality for
> the listener? I would expect so, but haven't been able to test. (All I
> know is that more often than not I'm using Skype to call people now
> (Skype-out service), because people are telling me they hear me
> better.)
>
> Note that you can very well say "you're wrong, we don't need a
> low-pass, just use a higher sampling rate then it sounds just fine",
> which is of course partially true, with higher sampling rates the
> relevance of a low-pass decreases quickly as the spectrum of
> frequencies that are higher than half the sample rate becomes narrower
> (and an increasing part of the consonant frequency range is digitized
> correctly instead of as noise). But given that at least Diamondcard
> only offers 8000 Hz, and there may be valid reasons for this (like
> maybe the phone system is still using that frequency), what's your
> position on this?

Isn't it waht the "plughw:0,0" special ALSA device is supposed to do by itself ?
We rely on it.

-- 
 _     Damien Sandras
(o-      
//\    Ekiga Softphone : http://www.ekiga.org/
v_/_   Be IP           : http://www.beip.be/
       FOSDEM          : http://www.fosdem.org/
       SIP Phone       : sip:dsandras ekiga net