Had a try. The reason why a simple delay is not that good is mainly due to the initialization of the filter parameter that still takes a few seconds (if they are perfectly in sync, you sort of get lucky). Otherwise, you real recording seems to have something odd in it. Are you sampling from a different card then the one that's playing the sound? or maybe the mic (or something else) in the room is moving while you're recording? Jean-Marc On Thu, 2005-11-10 at 07:28 -0800, Jason Harper wrote:> When I ran test 4 as originally described there is > substantial echo cancellation (but not as good as when > the files are perfectly aligned). When I invert the > inputs, there is no noticeable cancellation. > I'm using testecho with the preprocess line commented > out. Preprocess seems to work very well at cleaning > up the residual echo when mdf does its job, so I'm > just focusing on testing mdf. > > -Jason > > --- Jean-Marc Valin <Jean-Marc.Valin@USherbrooke.ca> > wrote: > > > First try your test #4 (same signal and offset) and > > swap the inputs. > > > > Jean-Marc > > > > Le mercredi 09 novembre 2005 ? 22:44 -0800, Jason > > Harper a ?crit : > > > I'm pretty much sure of it. When I test inverting > > the > > > inputs, my output is pretty much the same as my > > > speaker signal. Whereas the way that I normally > > test > > > the output is my mic signal with very little > > > attenuation. > > > If you are interested I can send my test files; > > they > > > are about 94KB each. > > > > > > -Jason > > > > > > --- Jean-Marc Valin > > <jean-marc.valin@usherbrooke.ca> > > > wrote: > > > > > > > Are you sure you're not just inverting the two > > > > inputs? > > > > > > > > Jean-Marc > > > > > > > > On Wed, 2005-11-09 at 22:16 -0800, Jason Harper > > > > wrote: > > > > > I ran some further tests on mdf and here are > > the > > > > > results: > > > > > 1. reduced tail length to 100ms, aligned mic > > and > > > > > speaker signals to within 10ms - almost no > > echo > > > > > attenuation > > > > > 2. aligned mic and speaker signals to within 5 > > > > samples > > > > > - still almost no echo attenuation > > > > > 3. ran testecho using the same file for mic > > and > > > > > speaker - very good echo cancellation (of > > course > > > > this > > > > > is expected, but I needed to do a sanity > > check) > > > > > 4. ran the same file for mic and speaker, but > > had > > > > the > > > > > mic signal delayed by 10ms - OK echo > > cancellation, > > > > but > > > > > noticeably degraded over previous test > > > > > 5. same file for mic and speaker, but mic > > signal > > > > > delayed by 40ms - noticeable echo attenuation, > > but > > > > > still a large echo. > > > > > > > > > > So far my testing is only showing good echo > > > > > cancellation when the exact same file is being > > > > used > > > > > and is pretty closely aligned. > > > > > > > > > > -Jason > > > > > > > > > > --- Jean-Marc Valin > > > > <Jean-Marc.Valin@USherbrooke.ca> > > > > > wrote: > > > > > > > > > > > This kind of behaviour is odd. One of the > > reason > > > > > > could be the fact that > > > > > > you're using a really long impulse response. > > Try > > > > > > syncing your signals > > > > > > and making the tail length more in the order > > of > > > > 100 > > > > > > ms to 300 ms. > > > > > > > > > > > > Jean-Marc > > > > > > > > > > > > Le dimanche 06 novembre 2005 ? 21:25 -0800, > > > > Jason > > > > > > Harper a ?crit : > > > > > > > Thanks for alerting me to the new changes. > > I > > > > just > > > > > > > tried the latest code from SVN, but > > > > unfortunately > > > > > > I > > > > > > > still have just about the same results. > > The > > > > > > estimated > > > > > > > echo that gets subtracted from the actual > > echo > > > > is > > > > > > such > > > > > > > a small signal that it doesn't really > > result > > > > in > > > > > > any > > > > > > > noticeable echo attenuation. > > > > > > > I currently have my filter size set to 2 > > > > seconds > > > > > > even > > > > > > > though the echo in my microphone file is > > only > > > > > > about > > > > > > > 180ms delayed from the signal in my > > speaker > > > > file. > > > > > > I > > > > > > > also double checked that the endianess of > > my > > > > is > > > > > > > correct, and that I am providing the files > > to > > > > > > testecho > > > > > > > in the correct order. > > > > > > > Do you think that there could be some > > > > > > characteristics > > > > > > > to my echo that make it not work with the > > mdf > > > > > > > algorithm? How much echo attenuation > > should I > > > > > > expect > > > > > > > from the mdf algorithm before the > > preprocess? > > > > Are > > > > > > > there any test speaker and mic files that > > I > > > > could > > > > > > run > > > > > > > against testecho to see how it handles > > other > > > > > > acoustic > > > > > > > echos? > > > > > > > > > > > > > > -Jason > > > > > > > > > > > > > > --- Jean-Marc Valin > > > > > > <Jean-Marc.Valin@USherbrooke.ca> > > > > > > > wrote: > > > > > > > > > > > > > > > I've recently made changes to the AEC. > > > > Please > > > > > > try > > > > > > > > the code in SVN and > > > > > > > > see if it works better. > > > > > > > > > > > > > > > > Jean-Marc > > > > > > > > > > > > > > > > Le jeudi 03 novembre 2005 ? 22:36 -0800, > > > > Jason > > > > > > > > Harper a ?crit : > > > > > > > > > I've tried some further debugging to > > see > > > > what > > > > > > mdf > > > > > > > > is > > > > > > > > > actually doing. Instead of sending: > > > > > > > > > tmp_out = (float)ref[i] - > > > > > > st->y[i+st->frame_size] > > > > > > > > > to the output, I just sent > > > > > > > > > st->y[i+st->frame_size] > > > > > > > > > to see what was being subtracted from > > the > > > > > > > > microphone > > > > > > > > > input. When I open this in Audacity, > > I > > > > see a > > > > > > very > > > > > > > > > small signal at about -40dBm. The > > actual > > > > echo > > > > > > in > > > > > > > > my > > > > > > > > > sample has a power closer to -20dBm. > > So, > > > > when > > > > > > the > > > > > > > > > -40dBm estimated echo is removed from > > the > > > > > > -20dBm > > > > > > > > > actual echo there is almost no > > perceived > > > > > > > > difference in > > > > > > > > > the echo power. > > > > > > > > > I'm sure that there are others on the > > list > > > > > > that > > > > > > > > have > > > > > > > > > had good success with the echo > > > > cancellation > > > > > > > > module. > > > > > > > > > Could anyone point me to reference > > test > > > === message truncated ===
Thanks for taking a look. There was no motion; however you are right about sampling from a different card. The speaker is connected to the Sound Blaster card, while the microphone is part of a USB webcam. I don't think that this is likely to be too unusual a configuration among users. I can retry the test using a sound card microphone to see if there is a difference. If it turns out that the problem is due to using different sound devices for speaker and microphone, does this mean that the mdf algorithm is particularly sensitive to the clock drift between devices? Is this because the frequencies of the two signals would be slightly off, reducing the correlation of the two signals? Any suggestions on how the algorithm could be tweaked to lessen this problem? -Jason --- Jean-Marc Valin <jean-marc.valin@usherbrooke.ca> wrote:> Had a try. The reason why a simple delay is not that > good is mainly due > to the initialization of the filter parameter that > still takes a few > seconds (if they are perfectly in sync, you sort of > get lucky). > Otherwise, you real recording seems to have > something odd in it. Are you > sampling from a different card then the one that's > playing the sound? or > maybe the mic (or something else) in the room is > moving while you're > recording? > > Jean-Marc > > > On Thu, 2005-11-10 at 07:28 -0800, Jason Harper > wrote: > > When I ran test 4 as originally described there is > > substantial echo cancellation (but not as good as > when > > the files are perfectly aligned). When I invert > the > > inputs, there is no noticeable cancellation. > > I'm using testecho with the preprocess line > commented > > out. Preprocess seems to work very well at > cleaning > > up the residual echo when mdf does its job, so I'm > > just focusing on testing mdf. > > > > -Jason > > > > --- Jean-Marc Valin > <Jean-Marc.Valin@USherbrooke.ca> > > wrote: > > > > > First try your test #4 (same signal and offset) > and > > > swap the inputs. > > > > > > Jean-Marc > > > > > > Le mercredi 09 novembre 2005 ? 22:44 -0800, > Jason > > > Harper a ?crit : > > > > I'm pretty much sure of it. When I test > inverting > > > the > > > > inputs, my output is pretty much the same as > my > > > > speaker signal. Whereas the way that I > normally > > > test > > > > the output is my mic signal with very little > > > > attenuation. > > > > If you are interested I can send my test > files; > > > they > > > > are about 94KB each. > > > > > > > > -Jason > > > > > > > > --- Jean-Marc Valin > > > <jean-marc.valin@usherbrooke.ca> > > > > wrote: > > > > > > > > > Are you sure you're not just inverting the > two > > > > > inputs? > > > > > > > > > > Jean-Marc > > > > > > > > > > On Wed, 2005-11-09 at 22:16 -0800, Jason > Harper > > > > > wrote: > > > > > > I ran some further tests on mdf and here > are > > > the > > > > > > results: > > > > > > 1. reduced tail length to 100ms, aligned > mic > > > and > > > > > > speaker signals to within 10ms - almost no > > > echo > > > > > > attenuation > > > > > > 2. aligned mic and speaker signals to > within 5 > > > > > samples > > > > > > - still almost no echo attenuation > > > > > > 3. ran testecho using the same file for > mic > > > and > > > > > > speaker - very good echo cancellation (of > > > course > > > > > this > > > > > > is expected, but I needed to do a sanity > > > check) > > > > > > 4. ran the same file for mic and speaker, > but > > > had > > > > > the > > > > > > mic signal delayed by 10ms - OK echo > > > cancellation, > > > > > but > > > > > > noticeably degraded over previous test > > > > > > 5. same file for mic and speaker, but mic > > > signal > > > > > > delayed by 40ms - noticeable echo > attenuation, > > > but > > > > > > still a large echo. > > > > > > > > > > > > So far my testing is only showing good > echo > > > > > > cancellation when the exact same file is > being > > > > > used > > > > > > and is pretty closely aligned. > > > > > > > > > > > > -Jason > > > > > > > > > > > > --- Jean-Marc Valin > > > > > <Jean-Marc.Valin@USherbrooke.ca> > > > > > > wrote: > > > > > > > > > > > > > This kind of behaviour is odd. One of > the > > > reason > > > > > > > could be the fact that > > > > > > > you're using a really long impulse > response. > > > Try > > > > > > > syncing your signals > > > > > > > and making the tail length more in the > order > > > of > > > > > 100 > > > > > > > ms to 300 ms. > > > > > > > > > > > > > > Jean-Marc > > > > > > > > > > > > > > Le dimanche 06 novembre 2005 ? 21:25 > -0800, > > > > > Jason > > > > > > > Harper a ?crit : > > > > > > > > Thanks for alerting me to the new > changes. > > > I > > > > > just > > > > > > > > tried the latest code from SVN, but > > > > > unfortunately > > > > > > > I > > > > > > > > still have just about the same > results. > > > The > > > > > > > estimated > > > > > > > > echo that gets subtracted from the > actual > > > echo > > > > > is > > > > > > > such > > > > > > > > a small signal that it doesn't really > > > result > > > > > in > > > > > > > any > > > > > > > > noticeable echo attenuation. > > > > > > > > I currently have my filter size set to > 2 > > > > > seconds > > > > > > > even > > > > > > > > though the echo in my microphone file > is > > > only > > > > > > > about > > > > > > > > 180ms delayed from the signal in my > > > speaker > > > > > file. > > > > > > > I > > > > > > > > also double checked that the endianess > of > > > my > > > > > is > > > > > > > > correct, and that I am providing the > files > > > to > > > > > > > testecho > > > > > > > > in the correct order. > > > > > > > > Do you think that there could be some > > > > > > > characteristics > > > > > > > > to my echo that make it not work with > the > > > mdf > > > > > > > > algorithm? How much echo attenuation > > > should I > > > > > > > expect > > > > > > > > from the mdf algorithm before the > > > preprocess? > > > > > Are > > > > > > > > there any test speaker and mic files > that >=== message truncated ===
To everyone on the list: do *NOT* attempt to do echo cancellation with signals sampled using different clocks. This will *NOT* work. Just a 0.1% difference between the two sampling rates (it's sometimes worse than that) means that the impulse response drifts by 8 samples every second. There's just no way to efficiently track this. Or at least no way that doesn't involve something 100x more complex than what I have now. So if you want to use two different soundcards, you should either: 1) Have professionals cards connected with a clock sync 2) Have an atomic clock in both cards 3) Forget about echo cancellation Jean-Marc Le jeudi 10 novembre 2005 ? 22:54 -0800, Jason Harper a ?crit :> Thanks for taking a look. There was no motion; > however you are right about sampling from a different > card. The speaker is connected to the Sound Blaster > card, while the microphone is part of a USB webcam. I > don't think that this is likely to be too unusual a > configuration among users. > I can retry the test using a sound card microphone to > see if there is a difference. If it turns out that > the problem is due to using different sound devices > for speaker and microphone, does this mean that the > mdf algorithm is particularly sensitive to the clock > drift between devices? Is this because the > frequencies of the two signals would be slightly off, > reducing the correlation of the two signals? Any > suggestions on how the algorithm could be tweaked to > lessen this problem? > > -Jason > > --- Jean-Marc Valin <jean-marc.valin@usherbrooke.ca> > wrote: > > > Had a try. The reason why a simple delay is not that > > good is mainly due > > to the initialization of the filter parameter that > > still takes a few > > seconds (if they are perfectly in sync, you sort of > > get lucky). > > Otherwise, you real recording seems to have > > something odd in it. Are you > > sampling from a different card then the one that's > > playing the sound? or > > maybe the mic (or something else) in the room is > > moving while you're > > recording? > > > > Jean-Marc > > > > > > On Thu, 2005-11-10 at 07:28 -0800, Jason Harper > > wrote: > > > When I ran test 4 as originally described there is > > > substantial echo cancellation (but not as good as > > when > > > the files are perfectly aligned). When I invert > > the > > > inputs, there is no noticeable cancellation. > > > I'm using testecho with the preprocess line > > commented > > > out. Preprocess seems to work very well at > > cleaning > > > up the residual echo when mdf does its job, so I'm > > > just focusing on testing mdf. > > > > > > -Jason > > > > > > --- Jean-Marc Valin > > <Jean-Marc.Valin@USherbrooke.ca> > > > wrote: > > > > > > > First try your test #4 (same signal and offset) > > and > > > > swap the inputs. > > > > > > > > Jean-Marc > > > > > > > > Le mercredi 09 novembre 2005 ? 22:44 -0800, > > Jason > > > > Harper a ?crit : > > > > > I'm pretty much sure of it. When I test > > inverting > > > > the > > > > > inputs, my output is pretty much the same as > > my > > > > > speaker signal. Whereas the way that I > > normally > > > > test > > > > > the output is my mic signal with very little > > > > > attenuation. > > > > > If you are interested I can send my test > > files; > > > > they > > > > > are about 94KB each. > > > > > > > > > > -Jason > > > > > > > > > > --- Jean-Marc Valin > > > > <jean-marc.valin@usherbrooke.ca> > > > > > wrote: > > > > > > > > > > > Are you sure you're not just inverting the > > two > > > > > > inputs? > > > > > > > > > > > > Jean-Marc > > > > > > > > > > > > On Wed, 2005-11-09 at 22:16 -0800, Jason > > Harper > > > > > > wrote: > > > > > > > I ran some further tests on mdf and here > > are > > > > the > > > > > > > results: > > > > > > > 1. reduced tail length to 100ms, aligned > > mic > > > > and > > > > > > > speaker signals to within 10ms - almost no > > > > echo > > > > > > > attenuation > > > > > > > 2. aligned mic and speaker signals to > > within 5 > > > > > > samples > > > > > > > - still almost no echo attenuation > > > > > > > 3. ran testecho using the same file for > > mic > > > > and > > > > > > > speaker - very good echo cancellation (of > > > > course > > > > > > this > > > > > > > is expected, but I needed to do a sanity > > > > check) > > > > > > > 4. ran the same file for mic and speaker, > > but > > > > had > > > > > > the > > > > > > > mic signal delayed by 10ms - OK echo > > > > cancellation, > > > > > > but > > > > > > > noticeably degraded over previous test > > > > > > > 5. same file for mic and speaker, but mic > > > > signal > > > > > > > delayed by 40ms - noticeable echo > > attenuation, > > > > but > > > > > > > still a large echo. > > > > > > > > > > > > > > So far my testing is only showing good > > echo > > > > > > > cancellation when the exact same file is > > being > > > > > > used > > > > > > > and is pretty closely aligned. > > > > > > > > > > > > > > -Jason > > > > > > > > > > > > > > --- Jean-Marc Valin > > > > > > <Jean-Marc.Valin@USherbrooke.ca> > > > > > > > wrote: > > > > > > > > > > > > > > > This kind of behaviour is odd. One of > > the > > > > reason > > > > > > > > could be the fact that > > > > > > > > you're using a really long impulse > > response. > > > > Try > > > > > > > > syncing your signals > > > > > > > > and making the tail length more in the > > order > > > > of > > > > > > 100 > > > > > > > > ms to 300 ms. > > > > > > > > > > > > > > > > Jean-Marc > > > > > > > > > > > > > > > > Le dimanche 06 novembre 2005 ? 21:25 > > -0800, > > > > > > Jason > > > > > > > > Harper a ?crit : > > > > > > > > > Thanks for alerting me to the new > > changes. > > > > I > > > > > > just > > > > > > > > > tried the latest code from SVN, but > > > > > > unfortunately > > > > > > > > I > > > > > > > > > still have just about the same > > results. > > > > The > > > > > > > > estimated > > > > > > > > > echo that gets subtracted from the > > actual > > > > echo > > > > > > is > > > > > > > > such > > > > > > > > > a small signal that it doesn't really > > > > result > > > > > > in > > > > > > > > any > > > > > > > > > noticeable echo attenuation. > > > > > > > > > I currently have my filter size set to > > 2 > > > > > > seconds > > > > > > > > even > > > > > > > > > though the echo in my microphone file > > is > > > > only > > > > > > > > about > > > > > > > > > 180ms delayed from the signal in my > > > > speaker > > > > > > file. > > > > > > > > I > > > > > > > > > also double checked that the endianess > > of > > > > my > > > > > > is > > > > > > > > > correct, and that I am providing the > > files > > > > to > > > > > > > > testecho > > > > > > > > > in the correct order. > > > > > > > > > Do you think that there could be some > > > > > > > > characteristics > > > > > > > > > to my echo that make it not work with > > the > > > > mdf > > > > > > > > > algorithm? How much echo attenuation > > > > should I > > > > > > > > expect > > > > > > > > > from the mdf algorithm before the > > > > preprocess? > > > > > > Are > > > > > > > > > there any test speaker and mic files > > that > > > === message truncated ==> >