wav2vec_new

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.2743
• Wer: 0.1376
• Cer: 0.1333

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
32.6273	1.0	26	32.5059	0.9394	0.7539
27.4795	2.0	52	22.9788	0.9996	0.9996
16.8834	3.0	78	7.6200	1.0	1.0
6.2939	4.0	104	4.0929	1.0	1.0
3.5881	5.0	130	3.4107	1.0	1.0
3.3449	6.0	156	3.2649	1.0	1.0
3.2703	7.0	182	3.2285	1.0	1.0
3.253	8.0	208	3.2188	1.0	1.0
3.2074	9.0	234	3.1777	1.0	1.0
3.2992	10.0	260	3.1773	1.0	1.0
3.2007	11.0	286	3.1827	1.0	1.0
3.1878	12.0	312	3.1493	1.0	1.0
3.1495	13.0	338	3.0784	1.0	1.0
3.082	14.0	364	2.9135	1.0	1.0
2.8619	15.0	390	2.7388	1.0	1.0
2.6435	16.0	416	2.4690	0.8928	0.9056
2.0885	17.0	442	1.7109	0.8447	0.8546
1.622	18.0	468	1.0533	0.4430	0.4414
0.8878	19.0	494	0.5311	0.2669	0.2651
0.4516	20.0	520	0.2743	0.1376	0.1333

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1