xls-r-asr_af-run5

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the asr_af dataset. It achieves the following results:

• Wer (Validation): 37.98%
• Wer (Test): 38.01%

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer (Train)
11.7451	0.58	50	7.2710	1.0
4.7931	1.17	100	4.0381	1.0
3.4944	1.75	150	3.1782	1.0
3.06	2.34	200	2.9951	1.0
3.0031	2.92	250	2.9964	1.0
2.9814	3.51	300	2.9652	1.0
2.9524	4.09	350	2.9419	0.9998
2.9014	4.68	400	2.8213	1.0
2.2569	5.26	450	1.6105	0.9433
1.3008	5.85	500	1.0090	0.8021
0.8965	6.43	550	0.7727	0.6551
0.7134	7.02	600	0.6579	0.6102
0.5275	7.6	650	0.5956	0.6005
0.4413	8.19	700	0.5558	0.5079
0.3582	8.77	750	0.5719	0.5459
0.296	9.36	800	0.5389	0.4822
0.2557	9.94	850	0.4608	0.4541
0.2051	10.53	900	0.4822	0.4290
0.1911	11.11	950	0.5035	0.4209
0.1635	11.7	1000	0.5319	0.4263
0.1582	12.28	1050	0.5075	0.4124
0.1387	12.87	1100	0.4759	0.4055
0.1251	13.45	1150	0.4925	0.3970
0.1164	14.04	1200	0.4933	0.3998
0.1052	14.62	1250	0.4587	0.3995
0.1023	15.2	1300	0.4863	0.3950
0.0918	15.79	1350	0.5114	0.3858
0.09	16.37	1400	0.5444	0.3940
0.086	16.96	1450	0.5071	0.3806
0.0798	17.54	1500	0.4914	0.3809
0.0728	18.13	1550	0.5425	0.3807
0.0678	18.71	1600	0.5221	0.3731
0.0667	19.3	1650	0.5239	0.3726

Framework versions

• Transformers 4.28.0
• Pytorch 2.0.1+cu118
• Datasets 2.14.4
• Tokenizers 0.13.3