Fig. 4
From: Free resources for forced phonetic alignment in Brazilian Portuguese based on Kaldi toolkit
Default architecture of the TDNN-F from Kaldi’s Mini-librispeech recipe. The setup is composed by multiple TDNN-F layers, although only one is being depicted. The dotted arrow represents a bypass connection similar to what happens in ResNet topologies. All layers but the output apply batch normalization, while \(l_2\) regularization is applied without exception. Linear blocks are intentionally thinner by either being bottlenecks or just having a lower dimension with respect to other layers