kul-speech-lab/whisper_medium_CGN

Whisper Medium CGN

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

• Loss: 0.2639
• Wer: 10.7278

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 64
• 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: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1116	1.01	1000	0.2978	15.2127
0.0786	2.03	2000	0.2842	13.4852
0.2042	3.04	3000	0.2656	13.3590
0.1183	4.05	4000	0.2667	12.6977
0.0584	6.01	5000	0.2604	12.0993
0.0126	7.02	6000	0.2776	12.1477
0.0837	8.04	7000	0.2541	11.9397
0.0229	9.05	8000	0.2663	11.3221
0.042	11.01	9000	0.2549	11.4863
0.0075	12.02	10000	0.2775	11.0780
0.008	13.03	11000	0.2499	10.9759
0.0739	14.05	12000	0.2308	10.9441
0.0379	16.01	13000	0.2423	10.7926
0.02	17.02	14000	0.2629	10.7699
0.0111	18.03	15000	0.2639	10.7278

Framework versions

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

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