w2v-bert-2.0-ln-afrivoice-10hr-v1

This model is a fine-tuned version of facebook/w2v-bert-2.0 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.3773
• Model Preparation Time: 0.0146
• Wer: 0.2579
• Cer: 0.0680

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
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Model Preparation Time	Wer	Cer
5.4983	1.0	31	2.8460	0.0146	1.0	1.0000
2.3085	2.0	62	1.1224	0.0146	0.7398	0.2589
0.8931	3.0	93	0.8311	0.0146	0.4138	0.1536
0.7396	4.0	124	0.7901	0.0146	0.7207	0.2228
0.6648	5.0	155	0.7516	0.0146	0.4619	0.1744
0.6398	6.0	186	0.7282	0.0146	0.4477	0.1611
0.586	7.0	217	0.8923	0.0146	0.3608	0.1344
0.5666	8.0	248	0.7070	0.0146	0.4131	0.1518
0.4841	9.0	279	0.7248	0.0146	0.3747	0.1437
0.4442	10.0	310	0.8866	0.0146	0.3733	0.1370
0.4325	11.0	341	0.6852	0.0146	0.3717	0.1429
0.3584	12.0	372	0.8831	0.0146	0.3453	0.1300
0.3397	13.0	403	0.9853	0.0146	0.3546	0.1314
0.3053	14.0	434	0.7321	0.0146	0.3924	0.1509
0.2655	15.0	465	0.8055	0.0146	0.3672	0.1386
0.2555	16.0	496	0.8419	0.0146	0.3725	0.1405
0.2239	17.0	527	0.8440	0.0146	0.3850	0.1449
0.1853	18.0	558	0.9243	0.0146	0.3681	0.1397
0.1554	19.0	589	0.9458	0.0146	0.3835	0.1452
0.125	20.0	620	1.2681	0.0146	0.3506	0.1309
0.1053	21.0	651	1.3723	0.0146	0.3682	0.1324
0.0776	22.0	682	1.3026	0.0146	0.3781	0.1369

Framework versions

• Transformers 4.44.0
• Pytorch 2.1.0+cu118
• Datasets 2.20.0
• Tokenizers 0.19.1