[codec] Hybrid codec simulations

[Jean-Marc Valin <[e-mail]>]

Hi everone,

There's been some discussion about the possibility of a hybrid codec, so 
I've made a few full-band codec experiments on the topic based on 
Broadvoice, SILK, and CELT. Note that these are simulations (based on 
*very* early work), not actual codecs. The idea is that a 
linear-prediction (LP) codec is used for the low band (0-4 Hz or 0-8 
kHz) and CELT is used up to 20 kHz.

I uploaded the files at http://jmvalin.ca/codec_experiment1/

The reference file is ref.wav. Here's an example of the naming I used: 
the file silk16_celt_40_21.wav means silk running at 16 kHz and 21 kbps 
and combined with CELT for a *total* bit-rate of 40 kb/s. For the BV 
files, the BV bit-rate is indicated directly with the codec.

Now, I should point out a few caveats (I'm sure there are more I forgot):

1) the filters aren't perfect reconstruction
2) the distribution of the bit-rate between the low-band and high-band 
was done very quicly
3) In the case of BV, I did not modify the codec, so the split is most 
likely not optimal
4) The CELT-only files probably have an unfair "advantage" here because 
these are the only files that have been tuned a bit because they are not 
a combination of two codecs
5) The CELT files (and combinations) are based on 10 ms frames, rather 
than the 5 ms frames I usually tune for (there's probably a few more 
birdie artefacts that there should be)
6) The SILK+CELT files are VBR, whereas the others are CBR
7) This is not meant to be a codec competition, just a way to explore 
the possibilities

All that being said, comments are welcome.

Cheers,

	Jean-Marc