metadata
language:
- en
tags:
- esc
datasets:
- librispeech
To reproduce this run, execute:
#!/usr/bin/env bash
python run_flax_speech_recognition_seq2seq.py \
--dataset_name="esc-benchmark/esc-datasets" \
--model_name_or_path="esc-benchmark/wav2vec2-aed-pretrained" \
--dataset_config_name="librispeech" \
--output_dir="./" \
--wandb_name="wav2vec2-aed-librispeech" \
--wandb_project="wav2vec2-aed" \
--per_device_train_batch_size="8" \
--per_device_eval_batch_size="2" \
--learning_rate="1e-4" \
--warmup_steps="500" \
--logging_steps="25" \
--max_steps="50001" \
--eval_steps="10000" \
--save_steps="10000" \
--generation_max_length="40" \
--generation_num_beams="1" \
--final_generation_max_length="300" \
--final_generation_num_beams="12" \
--generation_length_penalty="1.6" \
--hidden_dropout="0.2" \
--activation_dropout="0.2" \
--feat_proj_dropout="0.2" \
--overwrite_output_dir \
--gradient_checkpointing \
--freeze_feature_encoder \
--predict_with_generate \
--do_eval \
--do_train \
--do_predict \
--push_to_hub \
--use_auth_token