wavlm-base-ug-SALT

This model is a fine-tuned version of microsoft/wavlm-base on the AJIKADEV/SALT-MULTISPEAKER-ENG-SPLIT - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.2918
• Wer: 0.2407

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• 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: 500.0
• training_steps: 30000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3581	4.0	1000	0.3489	0.3778
0.1517	8.0	2000	0.3196	0.3228
0.1054	12.0	3000	0.3126	0.3190
0.0815	16.0	4000	0.3153	0.3174
0.0717	20.0	5000	0.3173	0.3005
0.0653	24.0	6000	0.3022	0.3013
0.0562	28.0	7000	0.3266	0.2967
0.0499	32.0	8000	0.3567	0.3111
0.0455	36.0	9000	0.3616	0.2914
0.041	40.0	10000	0.3337	0.2941
0.0364	44.0	11000	0.3307	0.2945
0.0329	48.0	12000	0.3231	0.2804
0.0285	52.0	13000	0.3193	0.2846
0.0269	56.0	14000	0.3407	0.2859
0.0225	60.0	15000	0.3022	0.2685
0.0212	64.0	16000	0.3226	0.2798
0.0184	68.0	17000	0.3173	0.2713
0.0168	72.0	18000	0.2998	0.2666
0.0144	76.0	19000	0.3141	0.2707
0.0127	80.0	20000	0.3115	0.2709
0.012	84.0	21000	0.3057	0.2662
0.0097	88.0	22000	0.3378	0.2611
0.0074	92.0	23000	0.3201	0.2603
0.0085	96.0	24000	0.3074	0.2584
0.0071	100.0	25000	0.2947	0.2529
0.0059	104.0	26000	0.2864	0.2538
0.0056	108.0	27000	0.2859	0.2511
0.0039	112.0	28000	0.3004	0.2466
0.0031	116.0	29000	0.2997	0.2431
0.0034	120.0	30000	0.2925	0.2412

Framework versions

• Transformers 5.0.0.dev0
• Pytorch 2.9.1+cu128
• Datasets 3.6.0
• Tokenizers 0.22.1