Vakyansh models in HuggingFace format

Installation

pip install -r requirements.txt
pip install https://github.com/kpu/kenlm/archive/master.zip

Usage

from infer import Wav2vecHF
model = Wav2vecHF(<path to models>)
model.transcribe(<path to wav file>)

Covert Vakyansh fairseq models to Huggingface format

cd wav2vec2_to_hf

Change the parameters in parameters.py appropriately:

HF_CONFIG = 'hf_config.json' #already present in repo
FINETUNED_CHECKPOINT = 'hindi_fairseq/him_4200.pt' # finetune checkpoint
PRETRAINED_CHECKPOINT = 'hindi_fairseq/CLSRIL-23.pt' 
DICT = 'hindi_fairseq/dict.ltr.txt'
LM = 'hindi_fairseq/lm.binary'
LEXICON = 'hindi_fairseq/lexicon.lst'
HF_DIRECTORY_PATH = 'hindi_hf'

python convert_checkpoint.py
python integrate_lm.py