whisper-malagasy-medium-v3

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

• Loss: 0.3918
• Wer: 0.2586

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3553	0.6510	1000	0.4508	0.3258
0.2512	0.6510	2000	0.4329	0.3045
0.2867	0.9766	3000	0.3841	0.2759
0.1985	1.3021	4000	0.3826	0.2826
0.2176	1.6276	5000	0.3660	0.2686
0.21	1.9531	6000	0.3558	0.2496
0.137	2.2786	7000	0.3668	0.2582
0.129	2.6042	8000	0.3670	0.2569
0.1373	2.9297	9000	0.3605	0.2474
0.0885	3.2552	10000	0.3772	0.2468
0.0891	3.5807	11000	0.3795	0.259
0.0956	3.9062	12000	0.3802	0.2499
0.0608	4.2318	13000	0.3918	0.2586

Framework versions

• Transformers 4.40.0
• Pytorch 2.4.1+cu124
• Datasets 2.19.0
• Tokenizers 0.19.1