Table 5 Results obtained with WRE approach compared to the baseline system. Digit Error Rate depending on the weight reestimation rules (MLE, FDW et FD) without ULT. 2 300 tests performed on BDSON corpus (clean).
From: Compact Acoustic Models for Embedded Speech Recognition
WRE | Baseline | |||
|---|---|---|---|---|
MLE | FDW | FD | ||
Very compact model | 3.35% | 2.78% | 3.13% | 4.96% |
Compact model | 2.83% | 2.17% | 2.48% | 4.32% |