python run_speech_recognition_ctc.py \
      --dataset_name="google/fleurs" \
      --dataset_config_name="ps_af" \
      --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
      --output_dir="./" \
      --overwrite_output_dir="False" \
      --max_steps="4500" \
      --per_device_train_batch_size="8" \
      --per_device_eval_batch_size="8" \
      --gradient_accumulation_steps="4" \
      --learning_rate="7.5e-5" \
      --warmup_steps="2000" \
      --evaluation_strategy="steps" \
      --text_column_name="transcription" \
      --save_steps="500" \
      --eval_steps="500" \
      --logging_steps="10" \
      --layerdrop="0.0" \
      --activation_dropout="0.1" \
      --eval_metrics wer cer \
      --greater_is_better="False" \
      --load_best_model_at_end \
      --save_total_limit="3" \
      --mask_time_prob="0.3" \
      --mask_time_length="10" \
      --mask_feature_prob="0.1" \
      --fp16 \
      --mask_feature_length="64" \
      --chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \
      --group_by_length \
      --push_to_hub \
      --do_train --do_eval \
      --gradient_checkpointing \
      --use_auth_token 
      --freeze_feature_extractor="True"