|
python run_speech_recognition_ctc.py \ |
|
--dataset_name "mozilla-foundation/common_voice_8_0" "collectivat/tv3_parla" "projecte-aina/parlament_parla" \ |
|
--dataset_config_name "ca" "ca" "clean" \ |
|
--model_name_or_path="facebook/wav2vec2-xls-r-300m" \ |
|
--train_split_name "train+validation" "train" "train+validation" \ |
|
--eval_split_name "test" "test" "test" \ |
|
--audio_column_name "audio" "audio" "audio" \ |
|
--output_dir="wav2vec2-xls-r-300m-ca" \ |
|
--overwrite_output_dir \ |
|
--num_train_epochs="10" \ |
|
--per_device_train_batch_size="32" \ |
|
--per_device_eval_batch_size="32" \ |
|
--gradient_accumulation_steps="4" \ |
|
--learning_rate="7.5e-5" \ |
|
--warmup_steps="2000" \ |
|
--length_column_name="input_length" \ |
|
--evaluation_strategy="steps" \ |
|
--text_column_name "sentence" "text" "sentence" \ |
|
--chars_to_ignore [ , ? . ! \; \: \" “ % ” � — … – ] \ |
|
--save_steps="500" \ |
|
--eval_steps="500" \ |
|
--logging_steps="500" \ |
|
--layerdrop="0.0" \ |
|
--activation_dropout="0.1" \ |
|
--save_total_limit="3" \ |
|
--freeze_feature_encoder \ |
|
--feat_proj_dropout="0.0" \ |
|
--mask_time_prob="0.75" \ |
|
--preprocessing_num_workers="12" \ |
|
--mask_time_length="10" \ |
|
--mask_feature_prob="0.25" \ |
|
--mask_feature_length="64" \ |
|
--gradient_checkpointing \ |
|
--use_auth_token \ |
|
--fp16 \ |
|
--group_by_length \ |
|
--do_train --do_eval \ |
|
--push_to_hub |
|
|
