Re: [Kaldi-users] Poll about Kaldi

[Tony R. <to...@ca...>]

Hi Dan,

I can contribute from both the research and the commercial perspectives.

I'll keep the research brief as that's been the main theme of 
contributions so far.   I think Kaldi should incorporate RNN acoustic 
models and a very efficient decoder (but I guess you could guess that 
I'd say this :-) .   That's our plan anyway - more detail is at 
http://cantabresearch.com/openings

 From a commercial perspective I think that the Kaldi scripts could 
really benefit from a refactoring.   At the moment it's set up from an 
academic viewpoint where you know all your data at the outset including 
the split into train and test - that's not the way things work for us.   
Vassil commented on this earlier.

* Kaldi code.   I think it would be cleaner if there was a single bin 
directory where all scripts and binaries were installed into (no more 
symlinks).   Running Kaldi would then be as simple as putting this 
directory into $PATH.

* Separate out data preparation, acoustic model generation,. decoder 
build and test.    In the commercial world you don't know the test set 
until it happens, so the same acoustic model will get used with many 
language models in many different deployments. Personally I like 
sourcing a config file which just has VARIABLE=VALUE definitions which 
sets everything up for the rest of the script.    Ideally there would be 
one set up and data preparation script and then the main acoustic 
train/decoder build/test would be three standard scripts that didn't 
need changing (the config file would tell you if you were going to stay 
at a GMM ML stage or progress to DNNs, etc).

* A stable release.   I deal with many people who want to switch from 
HTK to Kaldi, I encourage them to do it but not use kaldi-stable but to 
make their own.    An official stable release would lower the barrier to 
entry for many people, even if it's just called kaldi-2015H1 and 
kaldi-2015H2 and doesn't really change except very serious bugs (have 
there been any of those lately?  I think not).

If completed someone would be able to check out the latest stable 
release, run different sorts of models from kaldi-asr.org on different 
data, easily be able to adapt to different domains by changing the 
language model, etc.

Someone suggested to me that all this would be a good project for a 
Google Summer of Code student, I agree.


Tony

-- 
** Cantab is hiring: www.cantabResearch.com/openings **
Dr A J Robinson, Founder, Cantab Research Ltd
Phone direct: 01223 778240 office: 01223 794497
Company reg no GB 05697423, VAT reg no 925606030
51 Canterbury Street, Cambridge, CB4 3QG, UK