arampacha/wav2vec2-xls-r-1b-uk

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the /WORKSPACE/DATA/UK/COMPOSED_DATASET/ - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.1092
• Wer: 0.1752
• Cer: 0.0323

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: 5e-05
• train_batch_size: 16
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• training_steps: 12000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.7005	1.61	500	0.4082	0.5584	0.1164
1.1555	3.22	1000	0.2020	0.2953	0.0557
1.0927	4.82	1500	0.1708	0.2584	0.0480
1.0707	6.43	2000	0.1563	0.2405	0.0450
1.0728	8.04	2500	0.1620	0.2442	0.0463
1.0268	9.65	3000	0.1588	0.2378	0.0458
1.0328	11.25	3500	0.1466	0.2352	0.0442
1.0249	12.86	4000	0.1552	0.2341	0.0449
1.016	14.47	4500	0.1602	0.2435	0.0473
1.0164	16.08	5000	0.1491	0.2337	0.0444
0.9935	17.68	5500	0.1539	0.2373	0.0458
0.9626	19.29	6000	0.1458	0.2305	0.0434
0.9505	20.9	6500	0.1368	0.2157	0.0407
0.9389	22.51	7000	0.1437	0.2231	0.0426
0.9129	24.12	7500	0.1313	0.2076	0.0394
0.9118	25.72	8000	0.1292	0.2040	0.0384
0.8848	27.33	8500	0.1299	0.2028	0.0384
0.8667	28.94	9000	0.1228	0.1945	0.0367
0.8641	30.55	9500	0.1223	0.1939	0.0364
0.8516	32.15	10000	0.1184	0.1876	0.0349
0.8379	33.76	10500	0.1137	0.1821	0.0338
0.8235	35.37	11000	0.1127	0.1779	0.0331
0.8112	36.98	11500	0.1103	0.1766	0.0327
0.8069	38.59	12000	0.1092	0.1752	0.0323

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2
• Datasets 1.18.4.dev0
• Tokenizers 0.11.0