README.md

metadata

license: apache-2.0
datasets:
  - pongjin/en_corpora_parliament_processed
language:
  - en
pipeline_tag: automatic-speech-recognition
metrics:
  - wer

This model has been referred to the following links

1. https://huggingface.co/blog/wav2vec2-with-ngram
2. https://huggingface.co/blog/fine-tune-wav2vec2-english

Thanks to patrickvonplaten Patrick von Platen

해당 모델은 한국인의 영어 발화 인식 성능 개선을 위해 facebook/wav2vec2-base-960h 로 파인튜닝한 모델에 KenLM 5-gram 을 붙인 ASR + LM 모델입니다.

If you want to use LM, you must have kenlm installed https://github.com/kpu/kenlm

pip install https://github.com/kpu/kenlm/archive/master.zip

학습 데이터 출처 : https://aiopen.etri.re.kr/voiceModel

transformers==4.24.0
huggingface_hub==0.13.2

wer	epoch	batch	lr	weight_decay	warmup_steps
0.32	10	16	1e-4	0.005	1000