Figure 2.

The proposed marginalization-based method: A 2D feature sequence, $\mathit{F}=({\mathit{F}}_{1,1},\dots ,{\mathit{F}}_{H^{\prime },W^{\prime }})$, is produced by a 2D feature extractor such as a ViT backbone. $\mathit{F}$ is fed to a linear layer to produce $\mathit{S}=({\mathit{S}}_{1,1},\dots ,{\mathit{S}}_{H^{\prime },W^{\prime }})$ from which a softmax normalization is performed over both $H^{\prime }$ and C dimensions. Next, the normalized $\mathit{U}=({\mathit{U}}_{1,1},\dots ,{\mathit{U}}_{H^{\prime },W^{\prime }})$ is marginalized over the $H^{\prime }$ dimension to produce $\mathit{P}=({\mathit{P}}_1,\dots ,{\mathit{P}}_{W^{\prime }})$ that is fed to a CTC decoder. D and C are the feature and class dimensions, respectively.