| python run_phoneme_recognition_ctc.py \ | |
| --dataset_name "custom_datasets/aihub_broadcast_pr" "librispeech_asr" \ | |
| --model_name_or_path="facebook/wav2vec2-base" \ | |
| --dataset_config_name "clean" "clean" \ | |
| --train_split_name "train[:15%]" "train.100" \ | |
| --eval_split_name "validation[:8%]" "validation" \ | |
| --audio_column_name "audio" "audio" \ | |
| --text_column_name "text" "text" \ | |
| --phoneme_language "ko" "en-us" \ | |
| --output_dir="./saved/wav2vec2-base-en_kr_200h-pr" \ | |
| --cache_dir=".cache" \ | |
| --overwrite_output_dir \ | |
| --preprocessing_num_workers "64" \ | |
| --num_train_epochs="15" \ | |
| --eval_metrics "wer" "cer" \ | |
| --per_device_train_batch_size="32" \ | |
| --per_device_eval_batch_size="1" \ | |
| --gradient_accumulation_steps="1" \ | |
| --learning_rate="1e-4" \ | |
| --weight_decay="0.005" \ | |
| --warmup_steps="1000" \ | |
| --evaluation_strategy="steps" \ | |
| --save_steps="400" \ | |
| --eval_steps="100" \ | |
| --logging_steps="10" \ | |
| --layerdrop="0.0" \ | |
| --save_total_limit="3" \ | |
| --freeze_feature_encoder \ | |
| --ctc_zero_infinity \ | |
| --gradient_checkpointing \ | |
| --chars_to_ignore , ? . ! - \; \: \" \` “ % ‘ ” � "\b" \ | |
| --group_by_length \ | |
| --do_train --do_eval |