wavlm-salt-eng

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

• Loss: 0.2244
• Wer: 0.2118

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
0.3803	3.3228	1000	0.3107	0.3624
0.1663	6.6456	2000	0.2740	0.3097
0.1206	9.9684	3000	0.2486	0.2903
0.0938	13.2895	4000	0.2588	0.2828
0.0816	16.6123	5000	0.2769	0.2785
0.0689	19.9351	6000	0.2457	0.2882
0.0642	23.2562	7000	0.2639	0.2914
0.0566	26.5790	8000	0.2954	0.2828
0.049	29.9018	9000	0.3172	0.2763
0.0454	33.2230	10000	0.3186	0.2828
0.0395	36.5458	11000	0.2782	0.2817
0.0389	39.8686	12000	0.2857	0.2828
0.0321	43.1897	13000	0.2692	0.2527
0.0282	46.5125	14000	0.2570	0.2559
0.0269	49.8353	15000	0.2446	0.2624
0.0233	53.1564	16000	0.2383	0.2473
0.0224	56.4792	17000	0.2805	0.2473
0.0198	59.8020	18000	0.2555	0.2516
0.0159	63.1231	19000	0.2097	0.2409
0.015	66.4459	20000	0.2367	0.2505
0.015	69.7687	21000	0.2486	0.2527
0.0122	73.0899	22000	0.2475	0.2527
0.0104	76.4126	23000	0.2377	0.2344
0.008	79.7354	24000	0.2363	0.2441
0.0081	83.0566	25000	0.2347	0.2333
0.0072	86.3794	26000	0.2232	0.2290
0.0064	89.7022	27000	0.2212	0.2280
0.0044	93.0233	28000	0.2287	0.2258
0.004	96.3461	29000	0.2295	0.2344
0.0037	99.6689	30000	0.2243	0.2204

Framework versions

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