MSP-Audio

This model is a fine-tuned version of facebook/wav2vec2-large-robust-ft-libri-960h on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4829
• Wer: 0.2566
• Cer: 0.1474

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: 32
• eval_batch_size: 32
• seed: 42
• 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: cosine
• lr_scheduler_warmup_steps: 1000.0
• training_steps: 20000

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
1.5170	0.05	1000	0.2912	0.7151	0.4457
1.4106	0.1	2000	0.2405	0.5715	0.3834
1.3445	0.15	3000	0.2075	0.5755	0.3395
1.1670	0.2	4000	0.1713	0.4470	0.2948
1.1405	0.25	5000	0.1559	0.4444	0.2830
1.0518	0.3	6000	0.2054	0.6352	0.3497
1.0164	0.35	7000	0.1550	0.4675	0.2926
1.0954	0.4	8000	0.2192	0.6849	0.3549
1.0427	0.45	9000	0.1521	0.5033	0.2706
1.0515	0.5	10000	0.1804	0.6117	0.2952
0.9930	0.55	11000	0.1802	0.6416	0.2949
1.1711	0.05	12000	0.5594	0.2755	0.1603
1.0789	0.1	13000	0.4829	0.2566	0.1474
1.1322	0.15	14000	0.5620	0.2777	0.1640
0.9884	0.2	15000	0.4972	0.2594	0.1534
0.9589	0.25	16000	0.5521	0.2804	0.1689
0.9326	0.3	17000	0.5657	0.2834	0.1761
0.9061	0.35	18000	0.5497	0.2771	0.1701
0.9746	0.4	19000	0.5283	0.2681	0.1632
0.9603	0.45	20000	0.5331	0.2696	0.1639

Framework versions

• Transformers 5.10.2
• Pytorch 2.8.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2