elmamounedieye/ASR

Whisper Wolof Lengo AI V5

This model is a fine-tuned version of serge-wilson/whisper-small-wolof on the audiofolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.3569
• Wer: 36.0472
• Cer: 22.5967

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: 0.0005
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-05
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• training_steps: 1990
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.2672	1.0	208	1.2009	85.3838	63.2065
0.875	2.0	416	0.8801	95.6117	69.2841
0.5964	3.0	624	0.6979	88.4681	63.1476
0.3953	4.0	832	0.6112	69.2255	57.6000
0.2465	5.0	1040	0.5015	55.4825	44.1314
0.161	6.0	1248	0.4401	53.7476	36.3715
0.0903	7.0	1456	0.4081	47.1822	31.0320
0.0553	8.0	1664	0.3751	44.7783	29.2044
0.024	9.0	1872	0.3604	38.7686	25.2606
0.011	9.57	1990	0.3569	36.0472	22.5967

Framework versions

• Transformers 4.39.3
• Pytorch 2.2.2+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2