20base

This model is a fine-tuned version of facebook/wav2vec2-large-960h on the timit_asr dataset. It achieves the following results on the evaluation set:

• Loss: 0.3563
• Cer: 0.1174

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: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
6.1508	0.8	200	3.3750	0.9746
2.7527	1.61	400	1.1544	0.4298
1.1471	2.41	600	0.4961	0.1653
0.6961	3.21	800	0.4192	0.1432
0.6425	4.02	1000	0.4111	0.1366
0.5663	4.82	1200	0.3696	0.1319
0.5265	5.62	1400	0.3766	0.1345
0.4753	6.43	1600	0.3659	0.1350
0.4517	7.23	1800	0.3830	0.1320
0.4312	8.03	2000	0.3396	0.1286
0.4006	8.84	2200	0.3450	0.1234
0.3693	9.64	2400	0.3602	0.1279
0.3627	10.44	2600	0.3347	0.1223
0.3412	11.24	2800	0.3462	0.1271
0.3366	12.05	3000	0.3492	0.1227
0.3097	12.85	3200	0.3459	0.1242
0.2902	13.65	3400	0.3409	0.1189
0.2787	14.46	3600	0.3471	0.1194
0.2664	15.26	3800	0.3597	0.1192
0.2499	16.06	4000	0.3402	0.1173
0.2353	16.87	4200	0.3444	0.1174
0.2282	17.67	4400	0.3497	0.1185
0.2119	18.47	4600	0.3573	0.1192
0.207	19.28	4800	0.3563	0.1174

Framework versions

• Transformers 4.17.0
• Pytorch 2.4.0
• Datasets 1.18.3
• Tokenizers 0.20.3