whisper-small-af-ZA

This model is a fine-tuned version of openai/whisper-small on the google/fleurs dataset. It achieves the following results on the evaluation set:

• Loss: 0.5728
• Wer: 0.3664
• Wer Ortho: 0.3943

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: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 5
• training_steps: 2000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wer Ortho
0.7731	1.45	100	0.7280	0.3740	0.3863
0.2103	2.9	200	0.5116	0.3661	0.3859
0.0633	4.35	300	0.4967	0.2810	0.3008
0.0249	5.8	400	0.5003	0.3299	0.3477
0.0143	7.25	500	0.5191	0.3510	0.3660
0.0053	8.7	600	0.5149	0.3070	0.3221
0.0035	10.14	700	0.5345	0.3266	0.3443
0.0027	11.59	800	0.5339	0.3175	0.3344
0.0026	13.04	900	0.5435	0.3134	0.3328
0.0037	14.49	1000	0.5346	0.2506	0.2714
0.0045	15.94	1100	0.5438	0.3220	0.3389
0.0028	17.39	1200	0.5588	0.2551	0.2740
0.0036	18.84	1300	0.5466	0.2728	0.2702
0.0035	20.29	1400	0.5364	0.3119	0.3332
0.0056	21.74	1500	0.5608	0.2506	0.2721
0.0037	23.19	1600	0.5443	0.2833	0.3027
0.0035	24.64	1700	0.5466	0.3631	0.3866
0.0024	26.09	1800	0.5628	0.3198	0.3416
0.0036	27.54	1900	0.5495	0.2946	0.3122
0.0016	28.99	2000	0.5728	0.3664	0.3943

Framework versions

• Transformers 4.31.0.dev0
• Pytorch 1.12.1+cu116
• Datasets 2.4.0
• Tokenizers 0.12.1