wav2vec2-large-xls-r-300m-hungarian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HU dataset. It achieves the following results on the evaluation set:

• Loss: 0.2562
• Wer: 0.3112

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: 7e-05
• train_batch_size: 32
• eval_batch_size: 32
• 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: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.3964	3.52	1000	1.2251	0.8781
1.3176	7.04	2000	0.3872	0.4462
1.1999	10.56	3000	0.3244	0.3922
1.1633	14.08	4000	0.3014	0.3704
1.1132	17.61	5000	0.2913	0.3623
1.0888	21.13	6000	0.2864	0.3498
1.0487	24.65	7000	0.2821	0.3435
1.0431	28.17	8000	0.2739	0.3308
0.9896	31.69	9000	0.2629	0.3243
0.9839	35.21	10000	0.2806	0.3308
0.9586	38.73	11000	0.2650	0.3235
0.9501	42.25	12000	0.2585	0.3173
0.938	45.77	13000	0.2561	0.3117
0.921	49.3	14000	0.2559	0.3115

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0