python run_speech_recognition_ctc_for_cgn.py \
    --dataset_name="/project_antwerp/xlsr/data/cgn/cgn_dataset_16k" \
    --train_split_name="train" \
    --eval_split_name="test" \
    --text_column_name="text" \
    --audio_column_name="audio" \
    --length_column_name="input_length" \
    --min_duration_in_seconds="1.0" \
    --max_duration_in_seconds="20.0" \
    --model_name_or_path="./output.step1/" \
    --tokenizer_name_or_path="./output.step1/" \
    --output_dir="./output.step2" \
    --num_train_epochs="1" \
    --ignore_data_skip \
    --seed 43 \
    --per_device_train_batch_size="8" \
    --gradient_accumulation_steps="4" \
    --learning_rate="3e-5" \
    --warmup_steps="2000" \
    --evaluation_strategy="steps" \
    --save_steps="1000" \
    --eval_steps="3000" \
    --layerdrop="0.0" \
    --save_total_limit="4" \
    --freeze_feature_encoder \
    --gradient_checkpointing \
    --chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � _ \
    --fp16 \
    --group_by_length \
    --do_train --do_eval
    
#    --preprocessing_num_workers="4" \