sarpba/whisper-tiny-cleaned-128-6lr

Whisper Tiny Hu - cleaned

This model is a fine-tuned version of openai/whisper-tiny on the Common Voice 16.1 hu cleaned dataset. It achieves the following results on the evaluation set:

• Loss: 0.2208
• Wer Ortho: 21.4620
• Wer: 20.5992

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: 6e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 100
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
0.3889	0.32	1000	0.3614	38.3458	37.0630
0.2908	0.65	2000	0.2918	33.1232	31.7128
0.2699	0.97	3000	0.2543	28.6431	27.4796
0.1457	1.29	4000	0.2378	26.6342	25.4356
0.1449	1.62	5000	0.2306	25.6535	24.4806
0.1466	1.94	6000	0.2195	24.1181	23.1124
0.084	2.26	7000	0.2237	23.7987	22.6991
0.0823	2.59	8000	0.2188	22.8797	21.9703
0.0928	2.91	9000	0.2152	22.7368	21.6464
0.0561	3.23	10000	0.2158	22.1400	21.0376
0.0601	3.56	11000	0.2200	22.3698	21.3895
0.0616	3.88	12000	0.2153	20.9801	20.0073
0.0388	4.2	13000	0.2186	21.2182	20.1999
0.042	4.53	14000	0.2209	21.3611	20.3200
0.0444	4.85	15000	0.2208	21.4620	20.5992

Framework versions

• Transformers 4.37.2
• Pytorch 2.1.0+cu121
• Datasets 2.17.0
• Tokenizers 0.15.1