20add

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

• Loss: 0.3150
• Cer: 0.1185

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
5.7412	0.8	200	3.6319	0.8002
2.8636	1.61	400	1.4399	0.4842
0.8873	2.41	600	0.4543	0.1775
0.6085	3.21	800	0.5284	0.1427
0.5604	4.02	1000	0.4079	0.1434
0.5093	4.82	1200	0.7887	0.1367
0.4846	5.62	1400	0.3305	0.1336
0.4458	6.43	1600	0.3543	0.1305
0.4255	7.23	1800	0.3347	0.1326
0.4053	8.03	2000	0.3431	0.1303
0.395	8.84	2200	0.3165	0.1267
0.3749	9.64	2400	0.3284	0.1260
0.3731	10.44	2600	0.3404	0.1256
0.3464	11.24	2800	0.4583	0.1338
0.3347	12.05	3000	0.4427	0.1299
0.3201	12.85	3200	0.4347	0.1291
0.3141	13.65	3400	0.3283	0.1238
0.307	14.46	3600	0.3313	0.1235
0.298	15.26	3800	0.3079	0.1201
0.2858	16.06	4000	0.3074	0.1195
0.2764	16.87	4200	0.3108	0.1196
0.2776	17.67	4400	0.3164	0.1201
0.2632	18.47	4600	0.3187	0.1194
0.2609	19.28	4800	0.3150	0.1185

Framework versions

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