[Speex-dev] Speex 1.2beta1: Better, smaller, faster and more
jean-marc.valin at usherbrooke.ca
Wed Sep 6 16:00:00 PDT 2006
> I have been testing 1.2beta1 fixed point and non fixed point as well as 1.0.5
> compiled agains uclibc and run on 150Mhz pentium MMX chip.
> 1.2 non fixed is fastest 60 seconds encoded in 32 seconds at quality 4
> 1.0.5 is second with the same file encoding in 39 seconds at quality 4
> 1.2 fixed is slowest with 60 seconds encoded in 58 seconds at quality 4
> Is this as expected? Or am I doing something wrong? Can uclibc influence
The main thing it means is that the int performance of your CPU/compiler
isn't very good. I'm a *bit* surprised by how much difference there is,
so maybe it's a compiler issue as well. Of course, lowering the
complexity could help in all cases.
> We want to use speex and are considering a few ARM cpu's and the sis550 cpu
> (forerunner currently). Now I have to try and verify that it will encode from
> the soundcard on this device in realtime and transmit the data
> (multicasting). I was hoping that the non fixedpoint performance would be
> better, can it be my fault or is this expected behaviour/performance?
On an ARM chip, the fixed-point will definitely be faster because
there's no FPU.
> Is there a timeframe for the 1.2.0 stable release? Ballpark?
No. But if it makes you feel better, you can always try
% mv speex-1.2beta1.tar.gz speex-1.2.tar.gz
> Kobus Wolvaardt
> Kreon Technology
> On Monday 04 September 2006 16:09, Jean-Marc Valin wrote:
>> Speex 1.2beta1 announcement:
>> This new release brings many significant improvements. The quality has
>> been improved, both at the encoder level and the decoder level. These
>> include enhancer improvements (now on by default), input/output
>> high-pass filters, as well as fixing minor regressions in previous 1.1.x
>> releases. A strange and rare instability problem with pure sinusoids has
>> also been fixed. On top of that, memory use has been greatly reduced,
>> especially for fixed-point and narrowband. The fixed-point narrowband
>> encoder+decoder memory use has been cut by more than half, making it
>> possible to fit both in less than 6 kB of RAM. In general, CPU
>> requirement had gone down, especially for the fixed-point port. The
>> Blackfin port has been speeded up significantly, thanks to David Rowe.
>> There are also a few fixes for the TI C5X DSPs, as well as better
>> support for C++ compilers and crappy MS compilers. Oh, and before anyone
>> starts worrying, the format (bit-stream) itself has not changed, so
>> Speex is still compatible with version 1.0 and will continue to be in
>> the future.
>> Non-codec improvements include a extension (easier to use) to the echo
>> canceller API and a Speex-independent version of the jitter buffer. The
>> echo canceller should also be more robust to saturation in the capture
>> path. Last, but not least, the documentation has been updated.
>> Speex-dev mailing list
>> Speex-dev at xiph.org
More information about the Speex-dev