This folder contains configs and code related to the publication:

paper Dynamic Encoder Size Based on Data-Driven Layer-wise Pruning for Speech Recognition

We use RETURNN for training and our setups are based on Sisyphus.

We use models parts from i6-models

TED-LIUM-v2 Simple-Top-K

ConformerCTCModel, ConformerCTCConfig and train_step in returnn config is defined in here

TED-LIUM-v2 Iterative-Zero-Out

ConformerCTCModel, ConformerCTCConfig and train_step in returnn config is defined in here

LBS-960 Simple-Top-K

ConformerCTCModel, ConformerCTCConfig and train_step in returnn config is defined in here

LBS-960 Iterative-Zero-Out

ConformerCTCModel, ConformerCTCConfig and train_step in returnn config is defined in here

LBS-960 Aux-Loss

ConformerCTCModel, ConformerCTCConfig and train_step in returnn config is defined in here