[Cmusphinx-devel] LiveFeatureExtractor

[Philip K. <Philip.Kwok@Sun.COM>]

Hi,

I've been working making the Live decoder response faster. One of the
things in the front end that might potentially be slowing things down is
the DeltasFeatureExtractor (the one that computes delta-double-delta).
It slows things down because it processes incoming cepstra in batches of
25 frames. With 100 frames per second, this means a delay of 0.25 secs.
The delay is there because it was a straight port from s3.3.

I've worked on a version (LiveFeatureExtractor) that does not have such
a delay. The only delay it has is at the start of utterance, when it
will read ahead 3 frames. It replicates the first frame into 4 frames,
plus the next 3 frames (so you have {F0,F0,F0,F0,F1,F2,F3}), so that it
can compute the delta-double-delta on the first frame F0.

I haven't tested it live, so I'm not sure if it will work faster. In any
case, I think this should become the default FeatureExtractor, since the
code is much simpler (and better). It needs more testing, so I will
check it in, and test it on each set of regression tests (tidigits, an4,
rm1, etc..), one a day. If they all give the same WER, I will rename it
to DeltasFeatureExtractor, and retire the old one. Simple unit tests
show that it generates the same features as the old one.

Any objections or suggestions?

philip