[Kaldi-users] Realignment

[Daniel P. <dp...@gm...>]

Karel,
[and cc'ing kaldi-users because I think people will be interested in
this discussion],
You mentioned that for your BABEL recipe you tried re-aligning the
data with the neural net system and training again from scratch.

I tried this on a 100-hour subset of Switchboard and only got 0.1-0.2%
improvement; was this in the range of what you saw?

For reference, here is the old CPU-based recipe (it's slightly worse,
probably because I used more machines in parallel, 16, versus
8 in the GPU recipe):

%WER 19.0 | 1831 21395 | 83.1 11.5 5.4 2.1 19.0 57.6 |
exp/nnet5a_cpu/decode_eval2000_sw1_fsh_tgpr/score_11/eval2000.ctm.swbd.filt.sys
%WER 19.5 | 1831 21395 | 82.8 11.8 5.4 2.3 19.5 57.5 |
exp/nnet5a_cpu/decode_eval2000_sw1_tg/score_11/eval2000.ctm.swbd.filt.sys

Here is the GPU baseline:

%WER 18.0 | 1831 21395 | 83.7 10.7 5.5 1.7 18.0 55.5 |
exp/nnet5a_gpu/decode_eval2000_sw1_fsh_tgpr/score_12/eval2000.ctm.swbd.filt.sys
%WER 18.6 | 1831 21395 | 83.4 11.2 5.4 2.0 18.6 56.6 |
exp/nnet5a_gpu/decode_eval2000_sw1_tg/score_11/eval2000.ctm.swbd.filt.sys

And here is what I get when I realign the data with the system above,
and re-train the neural net from scratch using an otherwise similar
command:

%WER 17.9 | 1831 21395 | 83.8 10.8 5.5 1.7 17.9 55.8 |
exp/nnet6a_gpu/decode_eval2000_sw1_fsh_tgpr/score_13/eval2000.ctm.swbd.filt.sys
%WER 18.4 | 1831 21395 | 83.6 11.4 5.0 2.1 18.4 56.1 |
exp/nnet6a_gpu/decode_eval2000_sw1_tg/score_11/eval2000.ctm.swbd.filt.sys

Dan