8 Dec
2003
8 Dec
'03
6:09 a.m.
Hello,
I am looking for a deconvolver, that is able to produce impulse
responses from sinus sweeps (and especially the exponentially
sweeping sine wave introduced by Farina).
Do you have any suggestions or at least tips to start an
implementation by myself?
Recently I managed to use the mls tools from nwfiir to produce
an IR of my microverb.
I had to learn the hard way, that simple soundcards are not able
to be used as MLS source because of the non linearities. Even a
simple DA-AD loop gives a result wave that mls2imp cannot cope
with. But an empty loop with an US-122 (unfortunately not with
linux for now) gives something very near to a dirac impulse!
The hunt for the linux convolution reverb has started ;-)
Uwe
--
voiceINTERconnect www.voiceinterconnect.de
... smart speech applications from germany
8 Dec
8 Dec
1:54 p.m.
I would be interested in this project too. What I'm looking for is
actuallay an anti-reverb that will be able to cancel reverbs in a
listening room, well always in conjunction with the listener position.
The other useful thing would be a phase-filter to correct the phases
coming out from a two or three way loudspeaker to get clarity in the
sound similar to high-end speakers. I'm surprised that you mind modern
consumer soundcards not linear, after all the sigma-delta converters
used in most of todays soundcards are supposed to be perfectly linear
and it was one of the reasons of their adoption.
Apostolis
Uwe Koloska wrote:
Hello,
I am looking for a deconvolver, that is able to produce impulse
responses from sinus sweeps (and especially the exponentially sweeping
sine wave introduced by Farina).
Do you have any suggestions or at least tips to start an
implementation by myself?
Recently I managed to use the mls tools from nwfiir to produce an IR
of my microverb.
I had to learn the hard way, that simple soundcards are not able to be
used as MLS source because of the non linearities. Even a simple
DA-AD loop gives a result wave that mls2imp cannot cope with. But an
empty loop with an US-122 (unfortunately not with linux for now) gives
something very near to a dirac impulse!
The hunt for the linux convolution reverb has started ;-)
Uwe
9 Dec
9 Dec
5:20 a.m.
Hello,
Apostolos Dimitromanolakis wrote:
I would be interested in this project too. What
I'm looking
for is actuallay an anti-reverb that will be able to cancel
reverbs in a listening room, well always in conjunction with
the listener position.
Then I think DRC (digital room correction) is for you:
http://freshmeat.net/projects/drc/
This works with bruteFIR
http://www.ludd.luth.se/~torger/brutefir.html
But I must confess: I have not understand how to use the result
of drc in bruteFIR.
For a reverb convolution I think only the first steps of the
DRC process are necessary, but I haven't understand what to do
there, too.
And unfortunately the only DRC tutorial is for windows only
:-(( And it doesn't explain the steps, it just shows them ...
The other useful thing would be a phase-filter to
correct the
phases coming out from a two or three way loudspeaker to get
clarity in the sound similar to high-end speakers.
As far as I understand what DRC does, this is one of the
postprocessing steps.
I'm surprised that you mind modern consumer
soundcards not
linear, after all the sigma-delta converters used in most of
todays soundcards are supposed to be perfectly linear and it
was one of the reasons of their adoption.
Oh, I don't want to state, that all consumer soundcards are
nonlinear. I have a very old Soundblaster AWE 32 and found that
the waveform coming out when playing an mls signal cannot be
computed to an impulse response by mls2imp. (And looks very bad)
I hope to find some time to make a website showing the
waveforms and my experiences.
With a more modern USB AD/DA converter (Tascam US-122) I can
compute the IR both from a direct loop (what gives something
very near to a dirac) and from my alesis microverb.
And since the onboard sound of my ASUS motherboard with nvidia
nforce2 chipset gives similar results (though not so bad, but
still unusable), my only explanation (after examining the code
and process) was the unlinearity of the two soundcards. (But I
don't fully understand the whole process)
Maybe, if I post the waveforms, someone can give a better
explanation.
Another effect appears, when I fed the mls signal dircetly
through a reverb (ladspa, gverb). After the impulse, there is a
constant noise tail ...
Uwe
--
voiceINTERconnect www.voiceinterconnect.de
... smart speech applications from germany
10:41 a.m.
Hello,
At 10.17 09/12/2003 +0100, you wrote:
Exactly, it tries also to compensate for the room acoustic (part of, no
miracle can be performed here).
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
Hello,
Apostolos Dimitromanolakis wrote:
yep, you are right. To do the convolution between the measured sweep and
the inverse filter to get the impulse response you can use brutefir (with a
little trickery). I use it to do my own measurements with 45s log sweeps.
The inverse filter is almost 2 millions taps, but brutefir eat it without
esitation even with just 64 mb of RAM available. A truly wonderful piece of
software!
I would be interested in this project too. What
I'm looking
for is actuallay an anti-reverb that will be able to cancel
reverbs in a listening room, well always in conjunction with
the listener position.
Then I think DRC (digital room correction) is for you:
http://freshmeat.net/projects/drc/
This works with bruteFIR
http://www.ludd.luth.se/~torger/brutefir.html
But I must confess: I have not understand how to use the result
of drc in bruteFIR.
For a reverb convolution I think only the first steps of the
DRC process are necessary, but I haven't understand what to do
there, too. And unfortunately the only DRC tutorial is for
windows only
:-(( And it doesn't explain the steps, it just shows them ...
Doing it under Linux is a bit more complicated. If you want I have some
shell scripts that do all the steps needed to get the impulse response
(sweep playing & recording + deconvolution). Anyway they are just a clumsy
hack that I use myself for my measurements, don't expect any fancy
interface. Despite this, thanks to the brutefir floating point accuracy and
the long sweep used the results are state of the art (90+ dB of S/N even in
a not so quiet environment with a dirty cheap panasonic WM-60A capsule and
a DIY mic preamp).
The other
useful thing would be a phase-filter to correct the
phases coming out from a two or three way loudspeaker to get
clarity in the sound similar to high-end speakers.
As far as I understand what DRC does, this is one of the
postprocessing steps.
12:30 p.m.
Hello,
thank you for your answer!
Denis Sbragion wrote:
yep, you are right. To do the convolution between the
measured sweep and the inverse filter to get the impulse
response you can use brutefir (with a little trickery). I use
it to do my own measurements with 45s log sweeps. The
inverse filter is almost 2 millions taps, but brutefir eat it
without esitation even with just 64 mb of RAM available. A
truly wonderful piece of software!
But, how can I learn to use it in such a way??? I haven't found
any clue on the web page how to use a binary file as input data
for bruteFIR. Are your scripts helpful in that?
Doing it under Linux is a bit more complicated. If you
want I
have some shell scripts that do all the steps needed to get
the impulse response (sweep playing & recording +
econvolution). Anyway they are just a clumsy hack that I use
myself for my measurements, don't expect any fancy interface.
Despite this, thanks to the brutefir floating point accuracy
and the long sweep used the results are state of the art (90+
dB of S/N even in a not so quiet environment with a dirty
cheap panasonic WM-60A capsule and a DIY mic preamp).
Could you please send me these scripts. Since I am far more
better in UI design than in algortihmic debugging -- I hope to
make a nice tool from this skripts.
What do you use for creating the sweeps?
Let's make the linux convolution reverb real!
Yours
Uwe Koloska
--
voiceINTERconnect www.voiceinterconnect.de
... smart speech applications from germany
1:12 p.m.
Hello,
I already answered Uwe in his private e-mail. I copy also here my answer in
case it is of general help.
At 17.28 09/12/2003 +0100, you wrote:
But, how can I learn to use it in such a way??? I
haven't found
any clue on the web page how to use a binary file as input data
for bruteFIR. Are your scripts helpful in that?
no, this is in the BruteFIR documentation. BruteFIR accepts a lot of input
formats, including many binary formats, and the DRC output is, of course,
one of those supported. Look at the "Input and output structure" section of
the BF documentation:
http://www.ludd.luth.se/~torger/brutefir.html#config_4
there's the full list of supported formats. This list applies also to the
"Coeff structure":
http://www.ludd.luth.se/~torger/brutefir.html#config_3
which is the one that loads the filters. DRC, with the default
configurations files, outputs "FLOAT_LE" filters.
Could you please send me these scripts. Since I am far
more
better in UI design than in algortihmic debugging -- I hope to
make a nice tool from this skripts.
See the attached file. There are two examples for 44.1 Khz & 96 Khz
measurement (shell scripts + BF configuration). (Scripts not attached here,
of course. If anybody else is interested just write me in my private e-mail).
What do you use for creating the sweeps?
I used the Aurora plugins from Angelo Farina:
http://www.ramsete.com/aurora/
They're not free but they're free enough for my needs :) Anyway it's pretty
easy to code the log sweep procedure, I simply never found the time to do
it myself. If you want I can send you an AES paper (1.3 mb PDF) with
detailed explanations of the procedure and a complete comparison with other
measurements methods (log sweeps win hands off).
Let's make the linux convolution reverb real!
Aehm, to tell the true for me it's real since at least two years now :)
Let's change it to "let's make the Linux convolution reverb for the
masses". The convolution tools available for Linux are at the same time the
most sophisticated and the most difficult to use in the world. For sure DRC
is within the most difficult to use. May be it's because I developed it and
I'm also a Linux fan? :) AFAIK there are no more than a couple dozen people
in the whole world which are actually using it, and it's on the Internet
since almost one and a half year now.
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
11 Dec
11 Dec
5:20 p.m.
Am Dienstag, 9. Dezember 2003 18:06 schrieb Denis Sbragion:
Maybe I am too dumb, but I can't make it work ...
Here is what I have done:
1. create a sweep of 40s with AudioEase's "Make A TestTone 2.0"
(the wav files created with that prog cannot be read by sndfile-convert but
with wavesurfer/snack ...)
2. play the sweep with Spark LE on an iBook through a microverb into ardour
running on my linux box with NVIDIAs onboard sound
3. revert the sweep in time and create a float32 raw file for use with
bruteFIR
4. use bfcfg44 configuration for bruteFIR
but the output is empty (all floats have an exponent of -19..21)
And even the sweep directly fed into bruteFIR gives an empty file.
But I don't have any windows machine ;-)
Uwe
--
mailto:uwe@koloro.de
http://www.koloro.de
no, this is in the BruteFIR documentation. BruteFIR
accepts a lot of
input formats, including many binary formats,
uups, my fault -- I thought bruteFIR needs exponents of the FIR filter that
has to be calculated beforehand. But I don't realized that an impulse
response or a recorded sweep can be used directly.
Could you
please send me these scripts. Since I am far more
better in UI design than in algortihmic debugging -- I hope to
make a nice tool from this skripts.
See the attached file. There are two examples for 44.1 Khz & 96 Khz
measurement (shell scripts + BF configuration). What do you use
for creating the sweeps?
I used the Aurora plugins from Angelo Farina:
http://www.ramsete.com/aurora/
They're not free but they're free enough for my needs :)
12 Dec
12 Dec
7:30 p.m.
Am Donnerstag, 11. Dezember 2003 22:24 schrieb Uwe Koloska:
Maybe I am too dumb, but I can't make it work ...
At least for the convolution, I have managed to make it work.
The problem was with the attenuation. The values from Denis were not suited
for my setup. For the impulses I have around, the IR attenuation must be
24 to 32.
The deconvolution works, too. But I don't understand it completely. The IR
seems to start at the end of the file. Especially with the direct
deconvolution of the original IR, the pulse is at the end of the
convolution result ..
I must study the AES paper a bit deeper ;-)
Yours
Uwe
--
mailto:lad@koloro.de
http://www.koloro.de
15 Dec
15 Dec
5:20 a.m.
Hello Uwe,
At 00.28 13/12/2003 +0100, you wrote:
The deconvolution works, too. But I don't
understand it completely. The IR
seems to start at the end of the file. Especially with the direct
deconvolution of the original IR, the pulse is at the end of the
convolution result ..
it must be near the end of the file, because of the way it works.
Furthermore if you use BruteFIR to perform the deconvolution you should
consider that BruteFIR truncates the output to make it the same lenght of
the input, where the full convolution of a signal of length N with a filter
of length M would yeld a signal of length N+M-1. To avoid this truncation
simply add some seconds of silence to the recorded signal before performing
the convolution (the scripts I sent you already did it).
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
12:43 p.m.
Hello Denis,
just to get it right:
Denis Sbragion wrote:
it must be near the end of the file, because of the way it
works.
This is because of the delay of the filter line, isn't it?
Furthermore if you use BruteFIR to perform the
deconvolution
you should consider that BruteFIR truncates the output to
make it the same lenght of the input, where the full
convolution of a signal of length N with a filter of length M
would yeld a signal of length N+M-1. To avoid this truncation
simply add some seconds of silence to the recorded signal
before performing the convolution (the scripts I sent you
already did it).
So to get the IR right, I have to cut the beginning of the
bruteFIR result by the length of the filter (if the input starts
at the time the signal has started). Right?
How do you revise your IRs? Manually?
What is the reason for bruteFIR to keep this delay in the result?
Uwe
--
voiceINTERconnect www.voiceinterconnect.de
... smart speech applications from germany
7441
days inactive
7448
days old
linux-audio-dev@lists.linuxaudio.org
9 comments
4 participants
participants (4)
-
Apostolos Dimitromanolakis -
Denis Sbragion -
Uwe Koloska -
Uwe Koloska