[opus] Opus for ASR - update and questions

[Ross Bencina]

Hi Milan,

Did you try googling for "packet loss model"

I don't have a model for you but a few ideas. I worked on some packet 
delay models last year.. not exactly the same thing but similar.

In my experience, late packets are more prevalent than completely lost 
packets. Really late packets are as good as dropped.

So you could look at using some kind of queuing delay model and then 
threshold the delay time to determine which packets were effectively 
"dropped".

Another approach might be to use a uniform distribution of "drop events" 
-- a drop event might consist of a random number of sequentially dropped 
packets.

Or you could just run a bunch of real-world streams, capture the 
results, and use either the actual drop events, or the statistics of the 
tests as the basis of your tests or models.

I've observed a number of pathologies that are specific to particular 
routers, drivers and system setups (my home router puts huge 1 sec delay 
spikes in the traffic from time to time). WiFi packet scheduling can 
have some funky behavior too.. with respect to delays caused by channel 
scanning etc. So you would want to think about what parts of the system 
you're modelling (just copper/fibre network loss on a carrier-grade 
network? or also real-world domestic conditions).

I don't know how deeply you want to go into it but there's a field 
called Queuing Theory that will give you the mathematical tools to do 
the formal analysis. Kleinrock is the classic text I believe.

HTH

Ross.


On 7/12/2012 3:32 PM, Young, Milan wrote:
> Unfortunately, a uniform model is going to make the simulation behave
> better than it would in real-world busty conditions.  If anyone has
> alternate model suggestions, please let me know.  I need a good
> simulation and I'd be happy to donate code.
>
> Thanks
>
>> -----Original Message----- From: opus-bounces at xiph.org
>> [mailto:opus-bounces at xiph.org] On Behalf Of Timothy B. Terriberry
>> Sent: Thursday, December 06, 2012 6:46 PM To: opus at xiph.org
>> Subject: Re: [opus] Opus for ASR - update and questions
>>
>> Young, Milan wrote:
>>> One more follow-up question regarding the decoder's
>>> "packet-loss"
>> parameter.  What distribution model is being used to simulate
>> packet loss?  Is this a uniform model, or something that resembles
>> the bursty nature of real world conditions?
>>
>> The best answer here is to read the code (from opus_demo.c): lost =
>> len[toggle]==0 || (packet_loss_perc>0 && rand()%100 <
>> packet_loss_perc);
>>
>> So, uniform. _______________________________________________ opus
>> mailing list opus at xiph.org
>> http://lists.xiph.org/mailman/listinfo/opus
> _______________________________________________ opus mailing list
> opus at xiph.org http://lists.xiph.org/mailman/listinfo/opus
>