Pre-trained Conformer-CTC models for the librispeech dataset with icefall.
The model was trained on full LibriSpeech with the scripts in icefall.
See (https://github.com/k2-fsa/icefall/pull/13) for more details of this model.
How to use
See (https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/conformer_ctc/README.md)
Training procedure
The version of the mainly repositories are list below.
k2: https://github.com/k2-fsa/k2/commit/81cec9ec736d2c603ad75d933bb3e3a3706fb0dd
icefall: https://github.com/k2-fsa/icefall/commit/ef233486ae6d21bacb940de45efb35d0c334605c
lhotse: https://github.com/lhotse-speech/lhotse/commit/5dfe0f4c02b1334ebb7db6d67e1141fe406ca76b
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. It is better to use the given version above, but I think the latest version would be ok. And 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
git checkout ef233486
- Preparing data.
cd egs/librispeech/ASR
bash ./prepare.sh
- Training
export CUDA_VISIBLE_DEVICES="0,1,2,3"
python conformer_ctc/train.py --bucketing-sampler True \
--concatenate-cuts False \
--max-duration 200 \
--full-libri True \
--world-size 4
Evaluation results
The best decoding results (WERs) on LibriSpeech test-clean and test-other are listed below, we got this results by averaging models from epoch 15 to 34.
| test-clean | test-other | |
|---|---|---|
| WER | 2.57% | 5.94% |