scream_non_large_3e06_beams5_v2

This model is a fine-tuned version of openai/whisper-large-v2 on the NbAiLab/NCC_speech_all_v5 dataset. It achieves the following results on the evaluation set:

• step: 29999
• eval_loss: 1.2308
• train_loss: 0.2858
• eval_wer: 10.3228
• eval_cer: 5.6236

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-06
• lr_scheduler_type: linear
• per_device_train_batch_size: 5
• total_train_batch_size_per_node: 20
• total_train_batch_size: 160
• total_optimization_steps: 30,000
• starting_optimization_step: None
• finishing_optimization_step: 30,000
• num_train_dataset_workers: 32
• numb_hosts: 8
• total_num_training_examples: 4,800,000
• steps_per_epoch: 8543
• num_beams: 5

Training results

step	eval_loss	train_loss	eval_wer	eval_cer
0	2.0006	0.8864	15.7125	8.0675
5000	1.4781	0.3285	10.3532	5.0340
10000	1.2340	0.3277	10.8709	5.7596
15000	1.1556	0.3272	10.3837	5.3515
20000	1.2220	0.2775	10.2314	5.5026
25000	1.2454	0.2367	10.1705	5.4976
29999	1.2308	0.2858	10.3228	5.6236

Framework versions

• Transformers 4.29.0.dev0
• Datasets 2.11.0
• Tokenizers 0.13.3