Skip to content

Latest commit

 

History

History
20 lines (13 loc) · 1.02 KB

README.md

File metadata and controls

20 lines (13 loc) · 1.02 KB

Introduction

This recipe includes some different ASR models trained with WenetSpeech.

./RESULTS.md contains the latest results.

Transducers

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.