salmon-whisper-medium-smj

This model is a fine-tuned version of openai/whisper-medium on the NbAiLab/salmon-asr-smj dataset. It achieves the following results on the evaluation set:

• step: 9999
• validation_loss: 1.4491
• train_loss: 0.2105
• validation_wer: 15.6915
• validation_cer: 4.6710
• validation_exact_wer: 18.0851
• validation_exact_cer: 4.9990

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: 2.5e-05
• lr_scheduler_type: linear
• per_device_train_batch_size: 8
• total_train_batch_size_per_node: 64
• total_train_batch_size: 64
• total_optimization_steps: 10,000
• starting_optimization_step: None
• finishing_optimization_step: 10,000
• num_train_dataset_workers: 32
• num_hosts: 1
• total_num_training_examples: 640,000
• steps_per_epoch: 287
• num_beams: None
• weight_decay: 0.01
• adam_beta1: 0.9
• adam_beta2: 0.98
• adam_epsilon: 1e-06
• dropout: True
• bpe_dropout_probability: 0.2
• activation_dropout_probability: 0.1

Training results

step	validation_loss	train_loss	validation_wer	validation_cer	validation_exact_wer	validation_exact_cer
0	6.7175	4.8002	101.9947	49.5329	103.4574	50.6274
1000	1.8389	0.4886	22.2074	6.0723	26.1968	6.6321
2000	1.2170	0.2845	18.75	5.6661	22.8723	6.1741
3000	1.3422	0.3153	15.5585	4.7116	18.0851	5.0588
4000	1.2924	0.2716	16.4894	5.1178	20.7447	5.7558
5000	1.4563	0.2408	17.1543	5.1381	20.0798	5.5567
6000	1.4318	0.2653	16.2234	4.9147	19.4149	5.3774
7000	1.2483	0.2359	15.2926	4.8335	18.6170	5.2579
8000	1.4402	0.2701	15.4255	4.5695	18.2181	4.9592
9000	1.5363	0.2244	15.5585	4.6304	18.3511	5.0189
9999	1.4491	0.2105	15.6915	4.6710	18.0851	4.9990

Framework versions

• Transformers 4.35.0.dev0
• Datasets 2.14.6
• Tokenizers 0.14.1