xlsr-am-adap-phon

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 2.5869
• Wer: 0.9302
• Cer: 0.4393

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
6.6235	6.8966	100	6.3103	1.0	1.0
4.227	13.7931	200	4.2662	1.0	1.0
4.1461	20.6897	300	4.1543	1.0	0.9966
4.146	27.5862	400	4.1716	1.0	0.9859
4.105	34.4828	500	4.1391	1.0	0.9740
3.5688	41.3793	600	3.6625	1.0	0.9749
1.5705	48.2759	700	2.2315	1.0029	0.5187
0.6683	55.1724	800	2.2517	0.9684	0.4595
0.577	62.0690	900	2.2995	0.9528	0.4413
0.3109	68.9655	1000	2.4239	0.9397	0.4575
0.2803	75.8621	1100	2.4491	0.9508	0.4474
0.2136	82.7586	1200	2.4916	0.9179	0.4323
0.3282	89.6552	1300	2.5652	0.9302	0.4401
0.2118	96.5517	1400	2.5869	0.9302	0.4393

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1