icefall/egs/librispeech/ASR

Introduction

Please refer to https://icefall.readthedocs.io/en/latest/recipes/librispeech/index.html for how to run models in this recipe.

Transducers

There are various folders containing the name transducer in this folder. The following table lists the differences among them.

	Encoder	Decoder	Comment
transducer	Conformer	LSTM
transducer_stateless	Conformer	Embedding + Conv1d
transducer_lstm	LSTM	LSTM
transducer_stateless_multi_datasets	Conformer	Embedding + Conv1d	Using data from GigaSpeech as extra training data
pruned_transducer_stateless	Conformer	Embedding + Conv1d	Using k2 pruned RNN-T loss

The decoder in transducer_stateless is modified from the paper Rnn-Transducer with Stateless Prediction Network. We place an additional Conv1d layer right after the input embedding layer.