python run_speech_recognition_seq2seq_streaming.py \
    --output_dir="./" \
    --model_name_or_path="kimbochen/whisper-small-zh-tw" \
    --model_index_name="Whisper Small Traditional Chinese" \
    --dataset_name="mozilla-foundation/common_voice_11_0" \
    --dataset_config_name="zh-TW" \
    --language="chinese" \
    --train_split_name="train+validation" \
    --eval_split_name="test" \
    --learning_rate="1e-5" \
    --per_device_train_batch_size="64" \
    --per_device_eval_batch_size="32" \
    --max_steps="2000" \
    --warmup_steps="1200" \
    --logging_steps="25" \
    --evaluation_strategy="steps" \
    --eval_steps="400" \
    --save_strategy="steps" \
    --save_steps="400" \
    --generation_max_length="225" \
    --length_column_name="input_length" \
    --max_duration_in_seconds="30" \
    --text_column_name="sentence" \
    --freeze_feature_encoder="False" \
    --report_to="tensorboard" \
    --metric_for_best_model="wer" \
    --greater_is_better="False" \
    --load_best_model_at_end \
    --gradient_checkpointing \
    --fp16 \
    --overwrite_output_dir \
    --do_train \
    --do_eval \
    --predict_with_generate \
    --do_normalize_eval \
    --streaming \
    --use_auth_token \
    --push_to_hub