[Asterisk-Dev] IAX timing of VOICE frames

[Lee Howard]

Derek Smithies wrote:

>Hi,
> I found in the C++ version of IAX2 that I wrote for the OPAL library at 
>www.voxgratia.org that
>
>the timestamps had to represent the time at which the frame was 
>generated, not when it was actually sent.
>  
>

Thanks.  This helped quite a bit.  Apparently I was occassionally 
getting a short read from the original audio source, which threw off the 
"voice timestamp prediction" facility in the client.  (Some frames would 
actually end up with less than 160 samples, and that caused a shortened 
adjustment in the timestamp prediction.)  I fixed that by ensuring that 
all voice frames contained a full 160 samples (with the exception of 
possibly the last frame in a call).

However, I've also noticed that the time when the frame actually gets 
sent still influences things.  If I generate and send the frames exactly 
20000 u-sec apart I find frequent short (very short) audio cut-outs.  (I 
presume this is because of a "audio buffer underrun" of sorts either on 
the Asterisk server or on the other endpoint.)  If I generate and send 
the frames 19000 u-sec apart then I get "pops" in the audio.  (I presume 
this is because the server or the other endpoint has to "merge" a bit of 
audio from one frame with another because of the apparent clock skew, an 
"audio buffer overrun" - 20ms frames consistently arriving more quickly 
than every 20ms.)  The happy median seems to be around 19930 u-sec, but 
I suspect that would change if network conditions changed.

>Have you looked at the packet dump with ethereal? what do you observe - 
>are the packets coming too quickly, or are there some missing ??
>  
>

I have not looked at a packet dump with ethereal.  I'm not sure that I 
would know what to look for, anyway.

>What jitter buffer are you using ?
>Despite any urban myths you might have heard, a quality jitter buffer is
>essential to copy with variability in the rate at which packets are
>transmitted, and variability in network performance.
>  
>

I believe this entirely.

I'm using CVS v1_0, and I don't really understand what jitter buffer 
options there are.  I've tried with jitterbuffer=yes and with =no, in 
iax.conf, and I've tried enabling and disabling the buffer in the IAX 
client.

Currently I'm testing with a simple IAX client program written from a 
slightly modified version of libiax2 which is basically taking quickly 
repeating and predictable audio samples from a source and sending them 
through Asterisk to a SIP phone (Grandstream BT101) that I have.  I 
don't know if the appropriate jitter buffer to question here is that on 
the SIP phone or elsewhere.

Thanks,

Lee.