wav2vec2-large-xls-r-300m-finnish

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

• Loss: 0.2307
• Wer: 0.2984

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: 500
• num_epochs: 70.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9032	4.39	500	2.8768	1.0
1.5724	8.77	1000	0.5638	0.6438
1.1818	13.16	1500	0.3338	0.4759
1.0798	17.54	2000	0.2876	0.4086
1.0296	21.93	2500	0.2694	0.4248
1.0014	26.32	3000	0.2626	0.3733
0.9616	30.7	3500	0.2391	0.3294
0.9303	35.09	4000	0.2352	0.3218
0.9248	39.47	4500	0.2351	0.3207
0.8837	43.86	5000	0.2341	0.3103
0.8887	48.25	5500	0.2311	0.3115
0.8529	52.63	6000	0.2230	0.3001
0.8404	57.02	6500	0.2279	0.3054
0.8242	61.4	7000	0.2298	0.3006
0.8288	65.79	7500	0.2333	0.2997

Framework versions

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