wav2vec2-finetune-authentic-only

This model is a fine-tuned version of facebook/wav2vec2-base on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4832
• Wer: 0.3169

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: 0.0001
• train_batch_size: 32
• eval_batch_size: 8
• 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: 1000
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.2515	0.1337	500	2.9469	1.0
1.4619	0.2674	1000	0.9826	0.5937
0.7206	0.4011	1500	0.8086	0.4917
0.6012	0.5348	2000	0.7485	0.4530
0.5422	0.6684	2500	0.7128	0.4329
0.5063	0.8021	3000	0.6346	0.4055
0.479	0.9358	3500	0.6450	0.4001
0.44	1.0695	4000	0.6126	0.3856
0.4103	1.2032	4500	0.5970	0.3747
0.396	1.3369	5000	0.5792	0.3780
0.3822	1.4706	5500	0.5786	0.3643
0.3706	1.6043	6000	0.5387	0.3507
0.3669	1.7380	6500	0.5292	0.3546
0.3544	1.8717	7000	0.5145	0.3436
0.3492	2.0053	7500	0.5322	0.3342
0.3066	2.1390	8000	0.5284	0.3323
0.3006	2.2727	8500	0.5248	0.3333
0.2954	2.4064	9000	0.4983	0.3221
0.2914	2.5401	9500	0.4844	0.3202
0.2841	2.6738	10000	0.4881	0.3167
0.2816	2.8075	10500	0.4815	0.3167
0.2777	2.9412	11000	0.4832	0.3169

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0
• Datasets 3.5.0
• Tokenizers 0.21.1