Note: This recipe is trained with the codes from this PR https://github.com/k2-fsa/icefall/pull/233 And the SpecAugment codes from this PR https://github.com/lhotse-speech/lhotse/pull/604.

Pre-trained Transducer-Stateless models for the TEDLium3 dataset with icefall.

The model was trained on full TEDLium3 with the scripts in icefall.

Training procedure

The main repositories are list below, we will update the training and decoding scripts with the update of version.
k2: https://github.com/k2-fsa/k2 icefall: https://github.com/k2-fsa/icefall lhotse: https://github.com/lhotse-speech/lhotse

Install k2 and lhotse, k2 installation guide refers to https://k2.readthedocs.io/en/latest/installation/index.html, lhotse refers to https://lhotse.readthedocs.io/en/latest/getting-started.html#installation. I think the latest version would be ok. And please also install the requirements listed in icefall.
Clone icefall(https://github.com/k2-fsa/icefall) and check to the commit showed above.

git clone https://github.com/k2-fsa/icefall
cd icefall

Preparing data.

cd egs/tedlium3/ASR
bash ./prepare.sh

Training

export CUDA_VISIBLE_DEVICES="0,1,2,3"
./transducer_stateless/train.py \
                  --world-size 4 \
                  --num-epochs 30 \
                  --start-epoch 0 \
                  --exp-dir transducer_stateless/exp \
                  --max-duration 300

Evaluation results

The decoding results (WER%) on TEDLium3 (dev and test) are listed below, we got this result by averaging models from epoch 19 to 29. The WERs are

	dev	test	comment
greedy search	7.19	6.70	--epoch 29, --avg 11, --max-duration 100
beam search (beam size 4)	7.02	6.36	--epoch 29, --avg 11, --max-duration 100
modified beam search (beam size 4)	6.91	6.33	--epoch 29, --avg 11, --max-duration 100