[Speex-dev] Regression in wideband encoding quality between b1 and rc1

[Jean-Marc Valin]

On 16/12/09 06:12, Blaise Potard wrote:
> No. To tell you the truth, I tried yesterday to access the git
> repository to track the changes, but it seems to not be available at
> the moment, and I did not try any further. It still seems to be
> offline today, by the way.

We just moved a lot of services to a newer machine and it broke a few 
things. We're working on it.

>> 22 kHz is not supported. The encoder will give you a warning about that. You
>> can use the resampler to down-sample to 16 Hz.
>
> Yes, I know it is not officially supported, but it used to work quite
> well with beta 1 and sounded much better than 16000k

I guess that's what "not officially supported" means. If you can find 
the exact place where things broke *and* it's not just a tradeoff to 
make other things better, then I might fix it. No promise though.

> To be fair, there is a difference in quality also at 16k, but it is
> much harder to notice; I guess if you are maniac enough to be annoyed
> by it, you probably would not be using 16k to start with...

Did you check that the output rate is actually the same, because some of 
the VBR changes just lower both rate and quality at the same time, so 
you can easily just specify a higher quality level and obtain the same 
result as before.

> I gave it a quick try (thanks a lot for pointing it out, by the way!),
> and it definitely appears to be a really good alternative. We will
> most certainly switch to it in the future (but we will probably wait
> until version 1 is out).

Basically, as soon as you're willing to spend more than about 32-40 
kbps, CELT should almost always be better than Speex. Below that, stick 
to Speex.

	Jean-Marc