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