[Speex-dev] Struggling with AEC and OpenSL
roadriverrail at gmail.com
Tue Feb 4 13:30:43 PST 2014
Hi, Speex devs,
I apologize in advance if this is not the proper venue for this question,
but I had seen in the archives that other threads of this nature had been
In brief, I'm trying to get AEC working on a simple Android NDK app. It's
a basic "play from a file, record from the mic to file" loopback test. I'm
using OpenSL ES. I establish a player and a recorder at a 16kHz sampling
rate. OpenSL generally works by presenting buffers to players and
recorders, and you get called back when the buffer has been used. Both the
player and the recorder are operating with only one buffer, and in the
player and recorder callbacks from OpenSL, I make calls to Speex AEC.
Everything is working in a frame size of 160 samples (10ms of audio). The
source file is mono.
Obviously, I'm writing because I'm not able to ever cancel the echo. I
compiled Speex to dump the internal state of the echo canceler, and I found
a copy of echo_diagnostic.m online which works. Initially, echo_diagnostic
was telling me I had a far-end-to-near-end delay of 2700-3000 samples,
varying from test to test. So, I developed a buffering system to delay the
calls to speex_echo_playback(). With approximately 16 frames worth of
delay buffer, I have been able to drive the far-end-to-near-end delay down
to nearly 400 samples. Now, however, echo_diagnostic is warning me about
clock drift. It gives me messages like this:
"Drift estimate is -0.254065% (-160 samples)
Your clock is drifting! No way the AEC will be able to do anything with
that. Most likely, you're doing capture and playback from two different
I would think it's highly unlikely for a Google Nexus 7 tablet to have two
different sound cards for its speaker and its body mic, so I doubt that
this is the issue. I don't really have a good grasp of how the drift gets
measured and what, if anything, my own code could be doing to introduce it.
I'm feeling a bit frustrated because, while I understand what's
conceptually going on in AEC, I can't seem to make a simple echo test to
demonstrate it in implementation. I'd love any advice you'd be willing to
As an additional note, I based the native code for my test app off of the
code found here:
I'm not sure that code is right, either, because I have tried it out and
hear echo there, and echo_diagnostic measured the far-end-to-near-end delay
as -319 samples, despite the author's claims that the code works great and
I just need to adjust the tail length.
Thanks in advance for any advice any of you can offer me.
Rhett Aultman | http://about.me/rhettaultman
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Speex-dev