las-20s_asr-scr_w2v2-base_003

This model is a fine-tuned version of facebook/wav2vec2-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.7925
• Per: 0.1595
• Pcc: 0.6761
• Ctc Loss: 0.5196
• Mse Loss: 1.2099

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: 1e-05
• train_batch_size: 16
• eval_batch_size: 1
• seed: 3333
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2226
• training_steps: 22260
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Per	Pcc	Ctc Loss	Mse Loss
17.8143	3.0	2226	5.2572	0.9897	0.6607	3.7794	1.5450
4.3733	6.0	4452	4.7470	0.9719	0.6909	3.5988	1.3014
3.7753	9.0	6678	4.1218	0.9701	0.6900	3.1054	1.1960
2.3217	12.0	8904	2.1771	0.3106	0.6803	1.1828	1.0122
1.1884	15.0	11130	1.7753	0.2132	0.6855	0.7426	1.0054
0.8743	18.0	13356	2.0008	0.1827	0.6699	0.6203	1.3065
0.705	21.0	15582	2.0588	0.1705	0.6755	0.5638	1.4022
0.6121	24.0	17808	1.6289	0.1626	0.6753	0.5351	1.0556
0.5605	27.0	20034	1.7228	0.1605	0.6731	0.5245	1.1463
0.5226	30.0	22260	1.7925	0.1595	0.6761	0.5196	1.2099

Framework versions

• Transformers 4.38.1
• Pytorch 2.0.1
• Datasets 2.18.0
• Tokenizers 0.15.2