8 Jun
2006
8 Jun
'06
1:39 p.m.
Hi there.
I'm developing a little tool to automatically sync two audio
recordings of the same event made with two different recorders. I've
got the algorithms worked out and can correctly compute the drift and
offset between the files in my test recording. The next step is to
pad, stretch and mix the two files into one perfectly synced stereo
file.
The drift is very small. In my test recording the drift is only 53
milliseconds over the course of an hour. My stretch factor is
0.99998525762821
I've tried to stretch the file with sox:
sox infile.wav outfile.wav stretch 0.99998525762821
When I try to line up the files in a ardour I realise that the stretch
sox makes is hugely inaccurate.
For testing purposes I tried to stretch an hourly long file (44.1kHz)
by some 2000 samples, but the resulting file was some 30000 samples
longer. The stretch was off by more than 0.5 seconds!
Is there anything I can do to increase the accuracy of sox? Is there
any other tool out there that will do a better job at stretching with
very high accuracy?
alex
--
Alex Polite
http://flosspick.org - finding the right open source
8 Jun
8 Jun
1:59 p.m.
The drift is very small. In my test recording the
drift is only 53
milliseconds over the course of an hour. My stretch factor is
0.99998525762821
I'd tried to drop some samples in this case. Not by a large chunk, but
one by one periodically.
Yes, it's not correct but it may work out if you carefully choose when
to drop (for example, in silence).
5:07 p.m.
On 6/8/06, Dmitry Baikov <dsbaikov(a)gmail.com> wrote:
I'd tried to drop some samples in this case. Not
by a large chunk, but
one by one periodically.
Yes, it's not correct but it may work out if you carefully choose when
to drop (for example, in silence).
Yes, I've thought of this approach. What would be a good tool or
library to do that with? It would need to be something that can work
on the audio files without reading the whole files into memory first.
These files are very large.
alex
--
Alex Polite
http://flosspick.org - finding the right open source
7:15 p.m.
If the two files are both mono and you want to make them stereo then
judicious sample insertion will probably work reasonably well. If they are
both stereo then you need to look further into resampling since you may not
be able to insert samples at the same point in each of the respective
channels, and excessive insertion into one channel will introduce skew into
an otherwise synchronous pair of channels.
What is the format of the files? Raw audio, WAV, etc? What is the rough
content - if the signal is constant than sample insertion is not easy since
you may not have periods of silence. If the signal does have silent periods
then the solution may be a noise gate that inserts samples - this would be
all but inaudible. There is no reason why this should not work from file to
file with 'cat oldfile | <program> > newfile so should not need to have the
lot in memory.
Have you mailed the LAD list - somebody might be happy to bolt something
together for you, its not actually a lot of work for mono tracks.
Regards.
From: "Alex Polite"
<notmyprivateemail(a)gmail.com>
Reply-To: A list for linux audio users
<linux-audio-user(a)music.columbia.edu>
To: "A list for linux audio users" <linux-audio-user(a)music.columbia.edu>
Subject: [linux-audio-user] Stretching with very high accuracy
Date: Thu, 8 Jun 2006 17:39:48 +0200
Hi there.
I'm developing a little tool to automatically sync two audio
recordings of the same event made with two different recorders. I've
got the algorithms worked out and can correctly compute the drift and
offset between the files in my test recording. The next step is to
pad, stretch and mix the two files into one perfectly synced stereo
file.
The drift is very small. In my test recording the drift is only 53
milliseconds over the course of an hour. My stretch factor is
0.99998525762821
I've tried to stretch the file with sox:
sox infile.wav outfile.wav stretch 0.99998525762821
When I try to line up the files in a ardour I realise that the stretch
sox makes is hugely inaccurate.
For testing purposes I tried to stretch an hourly long file (44.1kHz)
by some 2000 samples, but the resulting file was some 30000 samples
longer. The stretch was off by more than 0.5 seconds!
Is there anything I can do to increase the accuracy of sox? Is there
any other tool out there that will do a better job at stretching with
very high accuracy?
alex
--
Alex Polite
http://flosspick.org - finding the right open source
_________________________________________________________________
Don't just search. Find. Check out the new MSN Search!
http://search.msn.click-url.com/go/onm00200636ave/direct/01/
7:42 p.m.
On Thu, Jun 08, 2006 at 11:15:03PM +0200, Nick Copeland wrote:
What is the format of the files? Raw audio, WAV, etc?
What is the rough
content - if the signal is constant than sample insertion is not easy since
you may not have periods of silence. If the signal does have silent periods
then the solution may be a noise gate that inserts samples - this would be
all but inaudible. There is no reason why this should not work from file to
file with 'cat oldfile | <program> > newfile so should not need to have the
lot in memory.
Given the very small amount of drift (53 ms / hour), we're talking about
inserting maybe 1 sample per 30000. I submit that nobody will notice
if you just insert a copy of an adjacent sample, and if you
want better than that, interpolation between two adjacent
samples would be even smoother.
--
Paul Winkler
http://www.slinkp.com
8:18 p.m.
From: Paul Winkler <pw_lists(a)slinkp.com>
Reply-To: A list for linux audio users
<linux-audio-user(a)music.columbia.edu>
To: linux-audio-user(a)music.columbia.edu
Subject: Re: [linux-audio-user] Stretching with very high accuracy
Date: Thu, 8 Jun 2006 17:42:36 -0400
On Thu, Jun 08, 2006 at 11:15:03PM +0200, Nick Copeland wrote:
You dont think that there may be issues with sample insertion at regular
locations? They can become noticable as they are cyclic - I tried it with
SLab, and for a while with bristol before deciding it needed resampling.
Perhaps going half way would be to duplicate a sample at zero crossing
rather than during silence. This would be a little more random and hence
less noticable.
Perhaps somebody should just fix sox - its such a cool tool that a fix might
be in order. Is it still maintained? I also thinks sox had a few different
options for resampling (binary/quadrature?), perhaps it just needs a
different option for the algorithm?
Nick
_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today it's FREE!
http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/
What is the format of the files? Raw audio, WAV,
etc? What is the rough
content - if the signal is constant than sample insertion is not easy
since
you may not have periods of silence. If the
signal does have silent
periods
then the solution may be a noise gate that
inserts samples - this would
be
all but inaudible. There is no reason why this
should not work from file
to
file with 'cat oldfile | <program> >
newfile so should not need to have
the
lot in memory.
Given the very small amount of drift (53 ms / hour), we're talking about
inserting maybe 1 sample per 30000. I submit that nobody will notice
if you just insert a copy of an adjacent sample, and if you
want better than that, interpolation between two adjacent
samples would be even smoother.
9 Jun
9 Jun
4:33 a.m.
On Thu, Jun 08, 2006 at 05:42:36PM -0400, Paul Winkler wrote:
Given the very small amount of drift (53 ms / hour),
we're talking about
inserting maybe 1 sample per 30000. I submit that nobody will notice
if you just insert a copy of an adjacent sample, and if you
want better than that, interpolation between two adjacent
samples would be even smoother.
I tried a more elaborate scheme some time ago (for Aeolus): instead
of just inserting on sample, replace N by N+1 samples using higher
order interpolation, with N up to 64. It could still be heard on some
sounds.
One way to resample accurately with ratios near to unity is to use
FFT based convolution to find the samples at fractional position
a/b and (a+1)/b, then interpolate linearly between those.
--
FA
Follie! Follie! Delirio vano e' questo!
10:46 a.m.
On 6/8/06, Nick Copeland <nickycopeland(a)hotmail.com> wrote:
What is the format of the files? Raw audio, WAV, etc?
What is the rough
content - if the signal is constant than sample insertion is not easy since
you may not have periods of silence. If the signal does have silent periods
then the solution may be a noise gate that inserts samples - this would be
all but inaudible. There is no reason why this should not work from file to
file with 'cat oldfile | <program> > newfile so should not need to have the
lot in memory.
Sounds cool. It also sounds simple enough for me to put together on my
own. The files are in mono wav. I could pipe them through sox to get
them to raw if that's easier to manipulate.
But I think I need to read up a bit on audio formats. I need to know
how many bytes to read to get one sample, given a specific sample
rate. And if I convert them to raw I need to find out how to get them
back to wav with the right header attached. What are the best places
to read up on that?
alex
--
Alex Polite
http://flosspick.org - finding the right open source
6551
days inactive
6552
days old
linux-audio-user@lists.linuxaudio.org
8 comments
5 participants
participants (5)
-
Alex Polite -
Dmitry Baikov -
Fons Adriaensen -
Nick Copeland -
Paul Winkler