wavlm-base-ug-demo

This model is a fine-tuned version of microsoft/wavlm-base on the AJIKADEV/UGANDA-COURT-SPEECH - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.1244
• Wer: 0.1379

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: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6536	0.8536	1000	0.4550	0.4195
0.3254	1.7068	2000	0.2806	0.2955
0.2162	2.5600	3000	0.2283	0.2417
0.1657	3.4131	4000	0.1805	0.2044
0.1344	4.2663	5000	0.1623	0.1951
0.1038	5.1195	6000	0.1489	0.1737
0.0833	5.9731	7000	0.1348	0.1626
0.0637	6.8263	8000	0.1310	0.1496
0.05	7.6795	9000	0.1272	0.1424
0.0409	8.5327	10000	0.1244	0.1379

Framework versions

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