xls-r-300m-pa

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - PA-IN dataset. It achieves the following results on the evaluation set:

• Loss: 1.0443
• Wer: 0.5715

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: 7.5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• 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: 2000
• num_epochs: 500.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.6694	19.22	500	4.0455	1.0
3.3907	38.45	1000	3.2836	1.0
2.0866	57.67	1500	1.2788	0.7715
1.4106	76.9	2000	0.7866	0.6891
1.1711	96.15	2500	0.6556	0.6272
1.038	115.37	3000	0.6195	0.5680
0.8989	134.6	3500	0.6563	0.5602
0.8021	153.82	4000	0.6644	0.5327
0.7161	173.07	4500	0.6844	0.5253
0.6449	192.3	5000	0.7018	0.5331
0.5659	211.52	5500	0.7451	0.5465
0.5118	230.75	6000	0.7857	0.5386
0.4385	249.97	6500	0.8062	0.5382
0.3984	269.22	7000	0.8316	0.5621
0.3666	288.45	7500	0.8736	0.5504
0.3256	307.67	8000	0.9133	0.5688
0.289	326.9	8500	0.9556	0.5684
0.2663	346.15	9000	0.9344	0.5708
0.2445	365.37	9500	0.9472	0.5590
0.2289	384.6	10000	0.9713	0.5672
0.2048	403.82	10500	0.9978	0.5762
0.1857	423.07	11000	1.0230	0.5798
0.1751	442.3	11500	1.0409	0.5755
0.1688	461.52	12000	1.0445	0.5727
0.1633	480.75	12500	1.0484	0.5739
0.1488	499.97	13000	1.0443	0.5715

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0