dataset API command -- kaggle competitions download -c bengaliai-speech

!python preprocess.py
--model_name_or_path "openai/whisper-small"
--language "Bengali"
--output_dir "75k-samples"
--preprocessing_num_workers 90
--preprocessing_only
--text_column_name "sentence"
--data_dir "data"
--min_duration_in_seconds 2
--max_duration_in_seconds 30
--max_train_samples 75000
--max_eval_samples 5000
--apply_spec_augment

#train.py data_dir = "/kaggle/input/bengali-ai-asr-10k"

!torchrun --nproc_per_node 2 train.py
--model_name_or_path "bangla-speech-processing/BanglaASR"
--train_data_dir $data_dir
--validation_data_dir $data_dir
--language "Bengali"
--output_dir "whisper-base-bn"
--do_train
--do_eval
--fp16
--group_by_length
--predict_with_generate
--dataloader_num_workers 1
--overwrite_output_dir
--per_device_train_batch_size 4
--length_column_name "input_length"
--report_to "none"
--metric_for_best_model "wer"
--greater_is_better False
--evaluation_strategy "epoch"
--save_strategy "epoch"
--save_total_limit 1
--logging_steps 10
--gradient_checkpointing
--warmup_steps 50
--apply_spec_augment True
--num_train_epochs 3
--learning_rate "1e-5"