[Speex-dev] Question from XM Radio

[Jim Crichton]

> On Sun, 2005-11-27 at 17:11 -0500, Carr, Terrance wrote:
>> Thanks for the reply.  We are currently using AMBE (4Kbps) for our
>> Traffic/Weather Channels.  If you have ever had a chance to hear the
>> service, you will know that AMBE does not do us well.
>>
>> I understand that 2Kbps is low quality, but any poorer than AMBE?
>
> If AMBE is 4 kbps, then most likely yes.

>
>> If can get a decent quality for other low bandwidth talk channels,
>> such as about 10-16Kbps and have it sound rather clean, then I would
>> be gaining a bit of growth in the areas I'm seeking.
>
> 10-16 kbps is different and Speex is probably more useful there.

AMBE has a pretty long history for low-bandwidth satellite speech (the 
predecessor IMBE is used in the Inmarsat-M system, and AMBE is used in the 
Inmarsat Mini-M and M4 at 3.6kbps and Iridium at 2.4 kbps).  A few years ago 
I did an evaluation of 4kbps speech codecs for a low-earth-orbit satellite 
operator (now defunct).  AMBE+ at 4kbps was intended to be a "toll quality" 
codec, with MOS scores similar to ACELP 8kbps.  It was clearly better than 
the couple of other 4kbps codecs which we tested, and MUCH better than 
LPC10.

Now, AMBE builds in forward error correction, which gives the most 
redundancy to the most important vectors in the encoded speech.  This is 
optimized for circuit mode systems, where all bits are delivered, even if 
some are in error.  This sort of FEC is useless in packet systems where the 
data is either delivered correct or not at all.  If XM is running speech in 
a packet mode, perhaps AMBE is not handing the loss of blocks of data data 
well.  This might not be the fault of the codec itself, but an issue with 
the way that it is integrated into the system.

Speex is unually flexible in its choice of bit and sampling rates, and you 
certainly have the ability to produce higher quality speech than you can get 
from AMBE.  But, as Jean-Marc points out, you will need a higher bitrate to 
do this.

- Jim