w2v2-lmk_updated

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

• Loss: 1.0281
• Wer: 0.5749
• Cer: 0.2049

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: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
9.4495	6.25	100	4.3525	1.0	1.0
3.1633	12.5	200	2.9339	1.0	1.0
2.964	18.75	300	2.8446	1.0	1.0
2.8489	25.0	400	2.6748	1.0	1.0
2.409	31.25	500	1.9297	1.0	0.6824
1.6654	37.5	600	1.3226	0.9024	0.3686
1.2589	43.75	700	1.0957	0.7666	0.2749
1.0072	50.0	800	1.0241	0.6585	0.2445
0.818	56.25	900	1.0104	0.6132	0.2239
0.7072	62.5	1000	0.9941	0.5923	0.2155
0.5998	68.75	1100	1.0503	0.5749	0.2079
0.5677	75.0	1200	1.0376	0.5784	0.2049
0.5125	81.25	1300	1.0549	0.5819	0.2133
0.4983	87.5	1400	1.0222	0.5505	0.1988
0.4859	93.75	1500	1.0176	0.5575	0.1995
0.4436	100.0	1600	1.0281	0.5749	0.2049

Framework versions

• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 3.0.0
• Tokenizers 0.22.1