Name		Name	Last commit message	Last commit date
parent directory ..
local		local
pruned_transducer_stateless2		pruned_transducer_stateless2
pruned_transducer_stateless5		pruned_transducer_stateless5
README.md		README.md
RESULTS.md		RESULTS.md
prepare.sh		prepare.sh
shared		shared

README.md

Introduction

This recipe includes some different ASR models trained with WenetSpeech.

./RESULTS.md contains the latest results.

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

	Encoder	Decoder	Comment
`pruned_transducer_stateless2`	Conformer(modified)	Embedding + Conv1d	Using k2 pruned RNN-T loss
`pruned_transducer_stateless5`	Conformer(modified)	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.