dmusingu/WHISPER-SMALL-SWAHILI-ASR-CV-14

Whisper small - Denis Musinguzi

This model is a fine-tuned version of openai/whisper-large on the Common Voice 14.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4428
• Wer: 0.2513
• Cer: 0.0983

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: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
0.9179	0.51	800	0.1412	0.5355	0.3693
0.3078	1.02	1600	0.1196	0.4343	0.3152
0.1959	1.53	2400	0.1172	0.4068	0.2822
0.1737	2.04	3200	0.1145	0.3922	0.2721
0.1046	2.55	4000	0.1084	0.3958	0.2634
0.1019	3.06	4800	0.1029	0.3957	0.2578
0.0588	3.57	5600	0.1132	0.4013	0.2666
0.0545	4.08	6400	0.1009	0.4112	0.2510
0.0305	4.59	7200	0.0941	0.4183	0.2442
0.0275	5.1	8000	0.1005	0.4303	0.2549
0.0153	5.61	8800	0.4374	0.2407	0.0908
0.014	6.12	9600	0.4428	0.2513	0.0983

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.1
• Datasets 2.17.0
• Tokenizers 0.15.2