deepspeech.pytorch.ko

A few changes has been made to train Korean. Please review deepspeech.pytorch in advance.

Koren ASR corpus

Korean read speech corpus (about 120 hours, 17GB) from National Institude of Korean Language

http://www.korean.go.kr/front/board/boardStandardView.do?board_id=4&mn_id=17&b_seq=464

https://ithub.korean.go.kr/user/corpus/referenceManager.do

Data preparation

cd data
python nikl.py
cd ..

Training

python train.py --train-manifest data/nikl_train_manifest.csv \
                --val-manifest data/nikl_val_manifest.csv \
                --cuda --save-folder nikl_models/

Prediction

python transcribe.py --model-path models/deepspeech_1.pth.tar \
                     --audio-path $HOME/NIKL/fv01/fv01_t01_s01.wav > trans.log
echo -e """$(cat trans.log | cut -d':' -f3 | cut -d'"' -f2)"""
> 직차도 점기도 없었다

python transcribe.py --model-path models/deepspeech_2.pth.tar \
                     --audio-path $HOME/NIKL/fv01/fv01_t01_s01.wav > trans.log
echo -e """$(cat trans.log | cut -d':' -f3 | cut -d'"' -f2)"""
> 기차도 전기도 없었다