icefall/egs/librispeech/ASR/transducer_stateless_multi_datasets

Introduction

The decoder, i.e., the prediction network, is from https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 (Rnn-Transducer with Stateless Prediction Network)

You can use the following command to start the training:

cd egs/librispeech/ASR
./prepare.sh
./prepare_giga_speech.sh

export CUDA_VISIBLE_DEVICES="0,1"

./transducer_stateless_multi_datasets/train.py \
  --world-size 2 \
  --num-epochs 60 \
  --start-epoch 0 \
  --exp-dir transducer_stateless_multi_datasets/exp-100 \
  --full-libri 0 \
  --max-duration 300 \
  --lr-factor 1 \
  --bpe-model data/lang_bpe_500/bpe.model \
  --modified-transducer-prob 0.25
  --giga-prob 0.2