Table 3 Parameters used for TDNN-F training
From: Free resources for forced phonetic alignment in Brazilian Portuguese based on Kaldi toolkit
Parameter | Value |
|---|---|
# TDNN-F layers | 12 |
# Epochs | 10 |
Time strides (TDNN-F layers) | {1, 1, 1, 0, 3, 3, 3, 3, 3, 3, 3, 3} |
Dimension | 768 on both TDNN-F and prefinal layers |
Bottleneck dimension | 96 on TDNN-F layers, 192 on prefinal layer |
Bypass scale | 0.66 |
Frame subsampling factor | 3 |
Regularization parameter | 0.015 for output layer, 0.03 otherwise |
Learning rate | 0.002 down to 0.0002 |