Pretrained model refuses to converge.

[MostafaAhmed98]

Hello,

Im trying to build an Arabic asr with common voice dataset with pretrained stt_en_citrinet_256 and after a 120 epochs i found that my model didn't improve and still making a bad WER (0.95) on test set.

i used sub-word sentencepiece tokenizer with 4096 size and changed the pretrained model vocab, tokenizer, train and test config as below:

from omegaconf import OmegaConf, open_dict
params = OmegaConf.load("./configs/config_bpe.yaml")
params.model.train_ds.manifest_filepath = 'train.json'
params.model.validation_ds.manifest_filepath = 'dev.json'
first_asr_model.change_vocabulary( new_tokenizer_dir='tokenizer_spe_unigram_v4096', new_tokenizer_type="bpe" )
first_asr_model.setup_training_data(train_data_config=params['model']['train_ds'])
first_asr_model.setup_validation_data(val_data_config=params['model']['validation_ds'])

below is my train, val loss and WER