Speex dev - Jun 2005 - trouble getting speex_echo_cancel() to work

> - set sampling rate to 8 kHz (at least for now)
> - make sure the far end signal in the playback signal is always a bit in
> advance (never late) compared to the mic signal.
> - Set the tail length to something around 100 ms.
>
> Also, if you're using two different soundcards (as I understand) for
the
> playback and the capture, you're *never* going to get echo canceller to
> work because the playback and capture need to be *exactly* in sync.
I did a bit more testing, and ended up creating a set of testsamples, as 
follows:

ideal.pcm => me saying "This is what I'd like to hear"
junk.pcm => me saying "With a bit of luck, this is gone"
mixdown.pcm => ideal.pcm, with junk.pcm started 7 ms later and 3 dB lower.

(All of this as 8000 Hz, 16-bit signed)

I then copied junk.pcm to play.sw, mixdown.pcm to mic.sw and ran testecho. 
The output sounds almost identical to the mixdown, and far from
"ideal".

As a 2nd test, I made play.sw and mic.sw identical (both the mixdown). 
The result is the same as the mixdown, with the volume slowly decaying 
until it's about 15 dB lower at the end of the 8 sec sample.

Somewhat encouraged by this, I made a much longer set of samples, and in 
the mixdown I inserted the "desired" sound 5 seconds into the mix to
let
the filter have time to adapt. Indeed, after a few seconds,  the echo is 
reduced by about 15 dB, but that will still be loud enough to be annoying 
and distractive when talking.

Are there any tunable parameters I can set to allow the echo canceller to 
be a bit more aggressive?

Hi,
 
Here http://home.arcor.de/andreadrian/ you can download another echo canceller
including a double talk detector. It is open source etc and it's used in
Kphone etc. I haven't got it to work well but this should be possible with
some tuning.
In my opinion it is easier to use than the echo canceller in speex
 
PS: the latest source of the echocanceller is included in the intercom program.
 
Greetings Jeroen de Kleijn

	-----Oorspronkelijk bericht----- 
	Van: speex-dev-bounces@xiph.org namens Thorvald Natvig 
	Verzonden: do 2-6-2005 15:58 
	Aan: speex-dev@xiph.org 
	CC: 
	Onderwerp: Re: [Speex-dev] trouble getting speex_echo_cancel() to work
	
	

	> - set sampling rate to 8 kHz (at least for now)
	> - make sure the far end signal in the playback signal is always a bit in
	> advance (never late) compared to the mic signal.
	> - Set the tail length to something around 100 ms.
	>
	> Also, if you're using two different soundcards (as I understand) for
the
	> playback and the capture, you're *never* going to get echo canceller
to
	> work because the playback and capture need to be *exactly* in sync.
	
	I did a bit more testing, and ended up creating a set of testsamples, as
	follows:
	
	ideal.pcm => me saying "This is what I'd like to hear"
	junk.pcm => me saying "With a bit of luck, this is gone"
	mixdown.pcm => ideal.pcm, with junk.pcm started 7 ms later and 3 dB lower.
	
	(All of this as 8000 Hz, 16-bit signed)
	
	I then copied junk.pcm to play.sw, mixdown.pcm to mic.sw and ran testecho.
	The output sounds almost identical to the mixdown, and far from
"ideal".
	
	As a 2nd test, I made play.sw and mic.sw identical (both the mixdown).
	The result is the same as the mixdown, with the volume slowly decaying
	until it's about 15 dB lower at the end of the 8 sec sample.
	
	Somewhat encouraged by this, I made a much longer set of samples, and in
	the mixdown I inserted the "desired" sound 5 seconds into the mix to
let
	the filter have time to adapt. Indeed, after a few seconds,  the echo is
	reduced by about 15 dB, but that will still be loud enough to be annoying
	and distractive when talking.
	
	Are there any tunable parameters I can set to allow the echo canceller to
	be a bit more aggressive?
	_______________________________________________
	Speex-dev mailing list
	Speex-dev@xiph.org
	http://lists.xiph.org/mailman/listinfo/speex-dev

> I did a bit more testing, and ended up creating a set of testsamples, as 
> follows:
> 
> ideal.pcm => me saying "This is what I'd like to hear"
> junk.pcm => me saying "With a bit of luck, this is gone"
> mixdown.pcm => ideal.pcm, with junk.pcm started 7 ms later and 3 dB
lower.
The problem with that is that it's nowhere near real conditions. It
would be a lot better if you made a real recording for mixdown because
then you would have a realistic echo function (with non-ideal freq
response, reverberation, ...).
> I then copied junk.pcm to play.sw, mixdown.pcm to mic.sw and ran testecho. 
> The output sounds almost identical to the mixdown, and far from
"ideal".
Are you using the latest code in svn (it has lots of improvements)?
> Somewhat encouraged by this, I made a much longer set of samples, and in 
> the mixdown I inserted the "desired" sound 5 seconds into the mix
to let
> the filter have time to adapt. Indeed, after a few seconds,  the echo is 
> reduced by about 15 dB, but that will still be loud enough to be annoying 
> and distractive when talking.
You also need the preprocessor hooks to get extra cancellation (see
manual.pdf in svn).
> Are there any tunable parameters I can set to allow the echo canceller to 
> be a bit more aggressive?
If you reduce the echo tail (filter size), it becomes more agressive
(otherwise you can tune the code yourself, but you may make it unstable
in the process).

	Jean-Marc

-- 
Jean-Marc Valin <Jean-Marc.Valin@USherbrooke.ca>
Universit? de Sherbrooke

>> I did a bit more testing, and ended up creating a set of testsamples,
as
>> follows:
>>
>> ideal.pcm => me saying "This is what I'd like to hear"
>> junk.pcm => me saying "With a bit of luck, this is gone"
>> mixdown.pcm => ideal.pcm, with junk.pcm started 7 ms later and 3 dB
lower.
>
> The problem with that is that it's nowhere near real conditions. It
> would be a lot better if you made a real recording for mixdown because
> then you would have a realistic echo function (with non-ideal freq
> response, reverberation, ...).
I thought I should give an "ideal" testcase to see it's best
theoretical
performance.
>> I then copied junk.pcm to play.sw, mixdown.pcm to mic.sw and ran
testecho.
>> The output sounds almost identical to the mixdown, and far from
"ideal".
>
> Are you using the latest code in svn (it has lots of improvements)?
No, I didn't, and having just upgraded, that made all the difference in 
the world. Doing the same test, the echo is now completely gone. That is, 
it's at least 30 dB lower than in the "echo" signal, which is low
enough
that it's not noticable when the person is talking, and also low enough 
not to trigger VAD during periods of silence.

It also handled the case where the echo (loudspeaker) signal was music 
which was 10 dB louder than the desired signal. The signal was noticably 
distorted from the original, but considering I couldn't even hear what was 
being said as long as the echo was there, it's very impressive.

Thank you very, very much :)