To reproduce this run, first call get_ctc_tokenizer.py to train the CTC tokenizer and then execute the following command to train the CTC system:

#!/usr/bin/env bash
python run_flax_speech_recognition_ctc.py \
 --model_name_or_path="esb/wav2vec2-ctc-pretrained" \
 --tokenizer_name="wav2vec2-ctc-spgispeech-tokenizer" \
 --dataset_name="esb/datasets" \
 --dataset_config_name="spgispeech" \
 --output_dir="./" \
 --wandb_project="wav2vec2-ctc" \
 --wandb_name="wav2vec2-ctc-spgispeech" \
 --max_steps="50000" \
 --save_steps="10000" \
 --eval_steps="10000" \
 --learning_rate="3e-4" \
 --logging_steps="25" \
 --warmup_steps="5000" \
 --preprocessing_num_workers="1" \
 --do_train \
 --do_eval \
 --do_predict \
 --overwrite_output_dir \
 --gradient_checkpointing \
 --freeze_feature_encoder \
 --push_to_hub \
 --use_auth_token