Table 7 Results obtained with ULT+WRE approach compared with the baseline. Digit Error Rate depending on the weight reestimation rules (MLE and FD) with ULT. 2 300 tests performed on BDSON corpus (clean).
From: Compact Acoustic Models for Embedded Speech Recognition
ULT+WRE | Baseline | ||
|---|---|---|---|
MLE | FD | ||
Very compact model | 3.04% | 3.39% | 4.96% |
Compact model | 2.78% | 2.26% | 4.32% |