metadata
language:
- en
tags:
- esb
datasets:
- esb/datasets
- speechcolab/gigaspeech
To reproduce this run, first install NVIDIA NeMo according to the official instructions, then execute:
#!/usr/bin/env bash
CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_rnnt.py \
--config_path="conf/conformer_transducer_bpe_xlarge.yaml" \
--model_name_or_path="stt_en_conformer_transducer_xlarge" \
--dataset_name="esb/datasets" \
--tokenizer_path="tokenizer" \
--vocab_size="1024" \
--num_train_epochs="0.88" \
--dataset_config_name="gigaspeech" \
--output_dir="./" \
--run_name="conformer-rnnt-gigaspeech" \
--wandb_project="rnnt" \
--per_device_train_batch_size="8" \
--per_device_eval_batch_size="4" \
--logging_steps="50" \
--learning_rate="1e-4" \
--warmup_steps="500" \
--save_strategy="steps" \
--save_steps="20000" \
--evaluation_strategy="steps" \
--eval_steps="20000" \
--report_to="wandb" \
--preprocessing_num_workers="4" \
--fused_batch_size="4" \
--length_column_name="input_lengths" \
--fuse_loss_wer \
--group_by_length \
--overwrite_output_dir \
--do_train \
--do_eval \
--do_predict \
--use_auth_token