opus - Aug 2015 - PLC Sounds Robotic - How to Implement FEC Wideband

I am specifically using Celt Wideband (48kHz) over WiFi multicast that naturally
leads to lost packets and am trying to minimize the impact to the audio.  I
implemented PLC but the audio it produces is robotic.  Have I implemented PLC
correctly?

Checking the waveform it is using the previous received waveform to fill in a
missing packet but not the full waveform so it has to repeat.   Basically,
within a single lost 20ms packet I hear two replicas of the last 10ms of the
previous waveform.  Sometimes the repeats are only 4ms long thus repeating 5
times.  This makes it sound robotic.  Is this typical or do I have it
implemented incorrectly?  Should it not try to create one replica of the entire
20ms previous waveform instead of having to repeat a portion of it?

Scott Boekweg
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/opus/attachments/20150825/28f2fe28/attachment.htm

What do you mean by "implement"? You're just using the Opus
built-in PLC
(passing NULL), right? The PLC generally attempts to find periodicity
and replicate it. I guess if your signal isn't periodic it can lead to a
repetition that isn't great. It's something that could probably be
improved in the PLC.

Cheers,

	Jean-Marc


On 08/25/2015 01:21 PM, Scott Boekweg wrote:
> I am specifically using Celt Wideband (48kHz) over WiFi multicast that
> naturally leads to lost packets and am trying to minimize the impact to
> the audio.  I implemented PLC but the audio it produces is robotic.
>  Have I implemented PLC correctly? 
> 
>  
> 
> Checking the waveform it is using the previous received waveform to fill
> in a missing packet but not the full waveform so it has to repeat.  
> Basically, within a single lost 20ms packet I hear two replicas of the
> last 10ms of the previous waveform.  Sometimes the repeats are only 4ms
> long thus repeating 5 times.  This makes it sound robotic.  Is this
> typical or do I have it implemented incorrectly?  Should it not try to
> create one replica of the entire 20ms previous waveform instead of
> having to repeat a portion of it?
> 
>  
> 
> Scott Boekweg
> 
> 
> 
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus
>

Thanks Jean-Marc for your prompt response.  Yes I am passing the Null so my
implementation is correct.  Your explanation of how it works helps explain what
is going on.  So it looks like FEC might be a better option.

After reading through Opus documentation it appears that in-band FEC is not
available for Celt, is this correct?  Assuming that's correct, is there any
recommended documentation on how to implement FEC with wideband Celt?  I'm
hoping I can pass the "redundant" FEC data through existing RTP
packets.

Many thanks!

Scott


-----Original Message-----
From: Jean-Marc Valin [mailto:jmvalin at jmvalin.ca] 
Sent: Tuesday, August 25, 2015 12:02 PM
To: Scott Boekweg <scott.boekweg at listentech.com>; opus at xiph.org
Subject: Re: [opus] PLC Sounds Robotic - How to Implement FEC Wideband

What do you mean by "implement"? You're just using the Opus
built-in PLC (passing NULL), right? The PLC generally attempts to find
periodicity and replicate it. I guess if your signal isn't periodic it can
lead to a repetition that isn't great. It's something that could
probably be improved in the PLC.

Cheers,

	Jean-Marc


On 08/25/2015 01:21 PM, Scott Boekweg wrote:
> I am specifically using Celt Wideband (48kHz) over WiFi multicast that 
> naturally leads to lost packets and am trying to minimize the impact 
> to the audio.  I implemented PLC but the audio it produces is robotic.
>  Have I implemented PLC correctly? 
> 
>  
> 
> Checking the waveform it is using the previous received waveform to 
> fill in a missing packet but not the full waveform so it has to repeat.
> Basically, within a single lost 20ms packet I hear two replicas of the 
> last 10ms of the previous waveform.  Sometimes the repeats are only 
> 4ms long thus repeating 5 times.  This makes it sound robotic.  Is 
> this typical or do I have it implemented incorrectly?  Should it not 
> try to create one replica of the entire 20ms previous waveform instead 
> of having to repeat a portion of it?
> 
>  
> 
> Scott Boekweg
> 
> 
> 
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus
>