Wav2vec 2.0

Foreigner Korean speech voice recognition hackathon - CSLEE

Requirements and Installation

• PyTorch version >= 1.10.0
• Python version >= 3.8
• To install foreignerspeech and develop locally:

git clone https://github.com/soohyunme/foreigner_speech
cd foreigner_speech
pip3 install -e .

• We only test this implementation in Ubuntu 20.04.
• DockerFile is also supported in this repo.

Instructions

• We support script examples to execute code easily(check script folder)

# Guide to make model with Foreigner-speech(orthographic transcription) 

# [1] unzip dataset
bash script/preprocess/unzip_foreigner_speech.sh

# [2] preprocess dataset & make manifest
bash script/preprocess/make_manifest.sh

# [3] further pre-train the model
bash script/pretrain/run_further_pretrain.sh
 
# [4] fine-tune the model
bash script/finetune/run_foreigner_speech.sh

# [5] inference the model
bash script/inference/evaluate_foreigner_speech.sh

Pretrained model

• E-Wav2vec 2.0 : Wav2vec 2.0 pretrained on Englsih dataset released by Fairseq(-py)
• Foreigner-wav2vec 2.0 : The model further pretrained on Foreigner-speech by using Englsih model
  • Fairseq Version : If you want to fine-tune your model with fairseq framework, you can download with this LINK

Dataset

• Foreigner-speech : Korean speech data of foreigner speakers

Acknowledgments

• Our code was modified from fairseq and K-wav2vec codebase. We use the same license as fairseq.
• The preprocessing code was developed with reference to Kospeech.

License

Our implementation code(-py) is MIT-licensed.