wav2vec2-xls-r-113m-id

This model is a fine-tuned version of evanarlian/distil-wav2vec2-xls-r-113m-id on the evanarlian/common_voice_11_0_id_filtered dataset. It achieves the following results on the evaluation set:

• Loss: 0.4804
• Wer: 0.4275

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: 0.0002
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 3
• total_train_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.3
• num_epochs: 30.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.2694	0.92	1000	2.9168	1.0000
2.2449	1.84	2000	1.5711	0.9901
1.2118	2.75	3000	1.0133	0.9261
0.971	3.67	4000	0.8860	0.8743
0.8472	4.59	5000	0.7562	0.8180
0.7436	5.51	6000	0.6800	0.7505
0.6603	6.43	7000	0.6275	0.7023
0.5961	7.35	8000	0.5913	0.6589
0.5458	8.26	9000	0.5605	0.6358
0.5113	9.18	10000	0.5346	0.6039
0.463	10.1	11000	0.5052	0.5689
0.4326	11.02	12000	0.4880	0.5497
0.3981	11.94	13000	0.4778	0.5357
0.3602	12.86	14000	0.4656	0.5198
0.3501	13.77	15000	0.4510	0.5085
0.3199	14.69	16000	0.4617	0.5010
0.3058	15.61	17000	0.4385	0.4880
0.2844	16.53	18000	0.4638	0.4930
0.2729	17.45	19000	0.4594	0.4783
0.2648	18.37	20000	0.4521	0.4703
0.2515	19.28	21000	0.4727	0.4627
0.2428	20.2	22000	0.4566	0.4587
0.2343	21.12	23000	0.4554	0.4545
0.2228	22.04	24000	0.4670	0.4506
0.2135	22.96	25000	0.4458	0.4446
0.2067	23.88	26000	0.4571	0.4402
0.2065	24.79	27000	0.4680	0.4359
0.1968	25.71	28000	0.4702	0.4346
0.1914	26.63	29000	0.4687	0.4320
0.182	27.55	30000	0.4807	0.4332
0.1771	28.47	31000	0.4824	0.4308
0.1728	29.38	32000	0.4804	0.4275

Framework versions

• Transformers 4.25.1
• Pytorch 1.12.1
• Datasets 2.7.1
• Tokenizers 0.13.1