wav2vec2-large-xls-r-300m-breton

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

• Loss: 0.6102
• Wer: 0.4455

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: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9205	3.33	500	2.8659	1.0
1.6403	6.67	1000	0.9440	0.7593
1.3483	10.0	1500	0.7580	0.6215
1.2255	13.33	2000	0.6851	0.5722
1.1139	16.67	2500	0.6409	0.5220
1.0688	20.0	3000	0.6245	0.5055
0.99	23.33	3500	0.6142	0.4874
0.9345	26.67	4000	0.5946	0.4829
0.9058	30.0	4500	0.6229	0.4704
0.8683	33.33	5000	0.6153	0.4666
0.8367	36.67	5500	0.5952	0.4542
0.8162	40.0	6000	0.6030	0.4541
0.8042	43.33	6500	0.5972	0.4485
0.7836	46.67	7000	0.6070	0.4497
0.7556	50.0	7500	0.6102	0.4455

Framework versions

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