phState
/ |
/ |
/ |
V V
PLP tandem
where tandem = [25-dim KLT'ed log activations from AF MLPs trained on Fisher, with SWB1 norms]
--
KarenLivescu - 03 Aug 2006
- training converged to 2%
- results on dev set
| WER | comp | lm_sc | lm_pen | notes |
| 58.7 | 128 | 16 | -1 | ckbeam 5000 |
| 60 | 128 | 16 | -1 | ckbeam 3000 |
--
ArtKantor - 07 Aug 2006
plp-mlp weight tuning:
| acc | gaussians | lm_scale | lm_penalty | plp weight |
| 40.0 | 128mix | 16 | -1 | 1 |
| 40.0 | 128mix | 16 | -1 | .7 |
| 39.5 | 128mix | 16 | -1 | .5 |
| 38.4 | 128mix | 16 | -1 | .3 |
| 38.3 | 128mix | 16 | -1 | 1.5 |
| 36.4 | 128mix | 16 | -1 | .1 |
The PLP weight of 1 is best
test on 2 percent converged: WER=61.3
test on .2 percent converged: WER=60.7
finalized
--
ArtKantor - 08 Aug 2006
train to .2 percent convergence, lm 16/-1
results on test set
| ckbeam | WER |
| 5000 | 60.7 |
| 10000 | 59.5 |
--
ArtKantor - 14 Aug 2006
