throatmic_subvocalization_whisper_base

This model is a fine-tuned version of openai/whisper-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.1228
• Wer: 0.5459

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: 5e-06
• train_batch_size: 16
• eval_batch_size: 64
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• training_steps: 800
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
7.8587	0.4464	25	6.9278	1.3758
5.3129	0.8929	50	3.6133	1.2219
2.3203	1.3393	75	1.8454	1.2380
1.4604	1.7857	100	1.4801	0.7549
1.2345	2.2321	125	1.3462	0.6779
1.0225	2.6786	150	1.2659	0.6481
0.8716	3.125	175	1.2188	0.6061
0.7413	3.5714	200	1.1807	0.6061
0.6884	4.0179	225	1.1523	0.5789
0.5424	4.4643	250	1.1423	0.5815
0.5548	4.9107	275	1.1269	0.5951
0.4597	5.3571	300	1.1245	0.5809
0.3878	5.8036	325	1.1131	0.5776
0.3648	6.25	350	1.1130	0.5459
0.3554	6.6964	375	1.1139	0.5621
0.294	7.1429	400	1.1131	0.5563
0.2431	7.5893	425	1.1235	0.5569
0.2668	8.0357	450	1.1108	0.5653
0.1998	8.4821	475	1.1228	0.5459

Framework versions

• Transformers 4.49.0
• Pytorch 2.6.0+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0