phState
/ | \
/ | ... \
V V |
dg1 pl1 |
| | |
| | |
V V |
VE_dg1 VE_pl1 PLPS
where VE_<F> is the virtual evidence given by MLP activations
for feature <F>. Dg1, pl1, etc. are *non*-deterministic given
the phone state. The PLPs are generated using mixtures of Gaussians.
Svitchboard, monophone, hybrid + PLPs
This system uses the 8 ANNs to provide virtual evidence about the 8 features. The 8 feature hidden RVs each depend on the phone state using a
DenseCPT
| Results for ANN outputs divided by the prior, trained using word alignments |
| Vocab size | Task | Word error rate (%) | VE scale factors | PLP scale factor | language model | Notes |
| | 1 | Validation | Test | dg1 | pl1 | | scale | penalty | |
| 10 | 22.1 | | 1.0 | 1.0 | 1.5 | 24 | -6 | Trained with all weights 1; weight tuning only during CV. CV running |
| 500 | | | | | | | | |
| | | | | | | | |
Validation means the D_short set, unless noted.
(1) Validation on only the first 100 utterances of D_short
See
HybridMonophoneNondetMapping for training recipes.
--
SimonKing - 07 Aug 2006
