[Freeswitch-users] Cepstral + FS question

[David Brazier]

On Tue, Dec 7, 2010 at 12:38 AM, Malay Thakershi <mthakershi at gmail.com> wrote:
> Hello, it would be great help if someone who has used Cepstral from FS can
> share their views.
> 1. I just have one Cepstral Allison voice license (1 port) on my FS server.
> I use swift command to convert text files to WAV which are then played by
> mod_managed in FS call process. First question is regarding limitations on
> simultaneous conversions (TXT to WAV) using swift command. If I have
> multiple threads doing this, will there be any degradation because of
> Cepstral? I tried running two BAT files with 3 commands each. But running
> them simultaneously or separate produced same outcome. Does anyone know when
> Cepstral licensing kicks in and starts degrading quality ( or worse
> inserting "not licensed") prompt?

The synthesis via the swift command is much faster than real time
(i.e. the time it takes to play the WAV) and is only restricted by CPU
speed and licenses.  You'd need to ask Cepstral the details, but my
simple tests have shown that it only uses one CPU per command though
some of the processing of multiple command is on multiple CPUs.  But I
think the licensing mechanism prevents multiple commands running
completely simultaneously on multiple CPUs.  I think the "not
licensed" speech is only inserted when you have no license at all -
all that happens if you try to run multiple commands is that they are
just delayed until  there is a free license.  I don't think quality
will ever be degraded - it's just a question of CPUs and licenses that
will determine your throughput.

> 2. When I call session speak from mod_managed (or stream file) after
> selecting Allison / Cepstral as my voice, does Cepstral engine interfere
> with quality of the playback? If yes, when will I see it and how can I
> produce their effects?

Not unless something in Cepstral or FS is going badly wrong!

> 3. What is the sensible number of ports (from Cepstral) I should be prepared
> to buy if findings in the previous points imposes significant limitations?

Depends on the length of your prompts and number and speed of CPUs and
anticipated simultaneous calls.  Try testing the time for a typical
prompt and work it out from there.

David