Reply by maury October 7, 20112011-10-07
On Oct 6, 2:59&#2013266080;pm, John McDermick <johnthedsp...@gmail.com> wrote:
> > John > > I got about 16 dB cancellation with an ordinary NLMS algorithm. What > > did you get, and what were you expecting? > > > Maurice > > You're a champ Maurice and really helpful with the feedback :-) Thank > you! > > I got about the same. What's bothering me is that the aec audio output > is 'smeared' > during double-talk and I'm not sure which parameters I should be > tweaking/changing > to make it sound better. It's the webRTC AEC...in case you want to > know... > > Maybe this smearing effect is just a limitation of AEC's ?? I have > seen cases where the > echo gets completely cancelled - even during doubletalk...so I'm > wondering what it is > that makes a difference...Is there an implicit limit to how powerful > the echo can be before > the near-end speech gets degraded/distorted with loss of > intelligibility...
What you're hearing is probably a poorly designed double-talk detector that doesn't stop adaptation before the impulse response estimate is affected. There is an art to double-talk detection design. Most are proprietary. Maurice
Reply by steveu October 7, 20112011-10-07
> >> >> John >> I got about 16 dB cancellation with an ordinary NLMS algorithm. What >> did you get, and what were you expecting? >> >> Maurice > >You're a champ Maurice and really helpful with the feedback :-) Thank >you! > >I got about the same. What's bothering me is that the aec audio output >is 'smeared' >during double-talk and I'm not sure which parameters I should be >tweaking/changing >to make it sound better. It's the webRTC AEC...in case you want to >know... > >Maybe this smearing effect is just a limitation of AEC's ?? I have >seen cases where the >echo gets completely cancelled - even during doubletalk...so I'm >wondering what it is >that makes a difference...Is there an implicit limit to how powerful >the echo can be before >the near-end speech gets degraded/distorted with loss of >intelligibility...
The WebRTC AEC is probably the exact thing you need. If you configure it correctly it has a feature to compensate for the slightly different sampling rates you get with the mic and speaker channels of a sound card. I think they actually implemented this to try to echo cancel over a VoIP link. This is a fruitless task in most cases, as you have little to no control over what codec translations, jitter buffering and other non-linear nasties may be going on in the channel. In your case you have just the kind of signals this AEC was designed for. I've been trying to find time to test out this canceler for myself, to solve the exact same problem you seem to be having. Steve
Reply by John McDermick October 6, 20112011-10-06
> John > I got about 16 dB cancellation with an ordinary NLMS algorithm. What > did you get, and what were you expecting? > > Maurice
any chance you could upload your aec output so I can compare and listen to how it sounds during double talk? Thanks
Reply by John McDermick October 6, 20112011-10-06
> > John > I got about 16 dB cancellation with an ordinary NLMS algorithm. What > did you get, and what were you expecting? > > Maurice
You're a champ Maurice and really helpful with the feedback :-) Thank you! I got about the same. What's bothering me is that the aec audio output is 'smeared' during double-talk and I'm not sure which parameters I should be tweaking/changing to make it sound better. It's the webRTC AEC...in case you want to know... Maybe this smearing effect is just a limitation of AEC's ?? I have seen cases where the echo gets completely cancelled - even during doubletalk...so I'm wondering what it is that makes a difference...Is there an implicit limit to how powerful the echo can be before the near-end speech gets degraded/distorted with loss of intelligibility...
Reply by maury October 6, 20112011-10-06
On Oct 6, 11:32=A0am, John McDermick <johnthedsp...@gmail.com> wrote:
> > Steve > > Yes, I remember. So I went on to confirm it by analyzing the signals > and that's > what I posted here. I just wanted to have somebody verify that my > analysis was > correct. > > Anyway, here are the audio files: > > http://www.nippyzip.com/uploads/111006113241-33140.zip
John I got about 16 dB cancellation with an ordinary NLMS algorithm. What did you get, and what were you expecting? Maurice
Reply by John McDermick October 6, 20112011-10-06
> > Steve
Yes, I remember. So I went on to confirm it by analyzing the signals and that's what I posted here. I just wanted to have somebody verify that my analysis was correct. Anyway, here are the audio files: http://www.nippyzip.com/uploads/111006113241-33140.zip
Reply by steveu October 6, 20112011-10-06
>Hello, > >I have made a script which slides a time-window over the speaker >signal and the microphone signal. For each window update I save the >time lag corresponding to the cross correlation peak value. I had >expected to see a more or less constant lag, but instead I see that it >is increasing over time. > >See chart here: > >http://www.nippyzip.com/uploads/111005105627-44838.zip > > >Wouldn't that indicate that the samplerate by which the microphone >signal is acquired is drifting? If not, what then?
Gee, you have a short memory. I gave you the answer a couple of days ago. It was an educated guess then. Now its confirmed. Steve
Reply by maury October 6, 20112011-10-06
On Oct 5, 9:57=A0am, John McDermick <johnthedsp...@gmail.com> wrote:
> Hello, > > I have made a script which slides a time-window over the speaker > signal and the microphone signal. For each window update I save the > time lag corresponding to the cross correlation peak value. I had > expected to see a more or less constant lag, but instead I see that it > is increasing over time. > > See chart here: > > http://www.nippyzip.com/uploads/111005105627-44838.zip > > Wouldn't that indicate that the samplerate by which the microphone > signal is acquired is drifting? If not, what then?
John, What are you using as the test signal? Can you upload your speaker and microphone files? Maurice Givens
Reply by John McDermick October 5, 20112011-10-05
Hello,

I have made a script which slides a time-window over the speaker
signal and the microphone signal. For each window update I save the
time lag corresponding to the cross correlation peak value. I had
expected to see a more or less constant lag, but instead I see that it
is increasing over time.

See chart here:

http://www.nippyzip.com/uploads/111005105627-44838.zip


Wouldn't that indicate that the samplerate by which the microphone
signal is acquired is drifting? If not, what then?