wav2vec2-large-xls-r-1b-frisian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice_13_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2204
• Wer: 0.1493

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: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.9606	2.45	300	2.6184	1.0
1.4992	4.9	600	0.4233	0.4143
0.9757	7.35	900	0.2765	0.3021
0.8773	9.8	1200	0.2529	0.2528
0.7448	12.24	1500	0.2363	0.2258
0.7039	14.69	1800	0.2258	0.2103
0.6811	17.14	2100	0.2217	0.2074
0.6279	19.59	2400	0.2050	0.1915
0.5938	22.04	2700	0.2229	0.1922
0.6227	24.49	3000	0.2088	0.2019
0.5682	26.94	3300	0.2127	0.1874
0.5939	29.39	3600	0.2044	0.1789
0.5427	31.84	3900	0.2185	0.1791
0.5551	34.41	4200	0.2097	0.1644
0.5021	36.86	4500	0.2180	0.1678
0.4589	39.31	4800	0.2076	0.1581
0.5204	41.76	5100	0.2181	0.1587
0.512	44.21	5400	0.2263	0.1607
0.465	46.66	5700	0.2204	0.1493

Framework versions

• Transformers 4.28.1
• Pytorch 2.0.0+cu117
• Datasets 2.11.0
• Tokenizers 0.13.3