CasperMuz/whisper-base-sna-cleaned

Whisper Medium Shona - Cleaned Data

This model is a fine-tuned version of openai/whisper-base on the Cleaned Google WAXAL Shona dataset. It achieves the following results on the evaluation set:

• Loss: 0.4745
• Wer: 38.1120

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: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use 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: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6465	0.5995	500	0.6614	53.6187
0.4528	1.1990	1000	0.4979	43.8992
0.4046	1.7986	1500	0.4480	40.3579
0.3495	2.3981	2000	0.4258	39.6383
0.3408	2.9976	2500	0.4087	37.3055
0.2667	3.5971	3000	0.4093	37.1480
0.2196	4.1966	3500	0.4092	36.8628
0.2352	4.7962	4000	0.4058	36.4691
0.1896	5.3957	4500	0.4145	36.3957
0.1973	5.9952	5000	0.4187	36.4935
0.1476	6.5947	5500	0.4271	37.2213
0.1244	7.1942	6000	0.4409	37.1643
0.1299	7.7938	6500	0.4430	36.8112
0.1107	8.3933	7000	0.4558	37.7590
0.1183	8.9928	7500	0.4572	37.9464
0.0944	9.5923	8000	0.4676	37.7400
0.0884	10.1918	8500	0.4707	37.9464
0.0873	10.7914	9000	0.4725	38.0605
0.0843	11.3909	9500	0.4742	38.0360
0.0836	11.9904	10000	0.4745	38.1120

Framework versions

• Transformers 5.14.1
• Pytorch 2.13.0+cu130
• Datasets 5.0.1
• Tokenizers 0.22.2