ICSI BN Feature Matrix
Matrix of 2000HU systems trained on 37h showing all combinations of rasta, per-utterance normalization and deltas, both alone and in combination with RNN1:
WERR%
|
alone
|
RNN combo
|
rasta
|
47.2
|
32.0
|
rasta-norm
|
45.7
|
32.1
|
plp
|
39.8
|
30.8
|
plp-norm
|
36.7
|
31.1
|
This table shows the overall word-error rate for the following conditions:
-
MLPs have 2000 hidden units and 54 output units
-
Nets have 13x9=117 input units for `plain' features, and 26x9=234 input units for features+deltas.
-
Training set is randomized half of 1996 set i.e. 37h total (same for all nets).
-
Test set is 32 minute subset of hub4 97 eval.
-
Noway pruning set to beam 2.0 state_beam 4.0 nhyps 7 prob_min 0.00020
The dimensions varied are:
-
Rasta filtering (rasta) vs. plain plp features (plp)
-
Per-utterance feature normalization (-norm) vs. no additional normalization
-
System used alone (alone) vs. system combined with Gary's best LNA system by averaging the log likelihoods (RNN combo)
-
Features alone used (WERR%) vs. features+deltas used (WERR%+deltas). Note that using deltas almost doubles the number of parameters in the net, but the training set size is kept the same.
Dan Ellis <dpwe@icsi.berkeley.edu>
Modified: $Date: 1998/06/22 06:16:02 $