kul-speech-lab/whisper_small_CGN

Whisper Small CGN

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

• Loss: 0.3386
• Wer: 15.1972

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1967	1.01	1000	0.4085	21.8459
0.1355	2.03	2000	0.3752	18.6212
0.2952	3.04	3000	0.3535	18.5841
0.1876	4.05	4000	0.3464	17.5097
0.1037	6.01	5000	0.3396	16.7360
0.0473	7.02	6000	0.3526	16.4131
0.1605	8.04	7000	0.3284	16.4012
0.0537	9.05	8000	0.3386	15.9454
0.0928	11.01	9000	0.3315	15.9568
0.0144	12.02	10000	0.3532	15.5387
0.0267	13.04	11000	0.3261	15.7577
0.0936	14.05	12000	0.3155	15.3380
0.0825	16.01	13000	0.3198	15.2653
0.0498	17.02	14000	0.3386	15.1972
0.0338	18.03	15000	0.3413	15.1972

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2

Whisper small model finetuned on Flemish part of Corpus Gesproken Nederlands (CGN).