whisper-medium-sb-lug-eng-v2

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

• Loss: 0.1255
• Wer Lug: 0.124
• Wer Eng: 0.171
• Wer Mean: 0.147
• Cer Lug: 0.028
• Cer Eng: 0.164
• Cer Mean: 0.096

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: 16
• eval_batch_size: 8
• 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Lug	Wer Eng	Wer Mean	Cer Lug	Cer Eng	Cer Mean
0.7222	0.1	500	0.2662	0.391	0.146	0.268	0.085	0.104	0.095
0.5452	0.2	1000	0.1894	0.223	0.058	0.141	0.05	0.034	0.042
0.4172	0.3	1500	0.1764	0.199	0.276	0.238	0.043	0.269	0.156
0.3949	0.4	2000	0.1583	0.177	0.035	0.106	0.039	0.017	0.028
0.3626	0.5	2500	0.1508	0.153	0.157	0.155	0.035	0.122	0.078
0.3467	0.6	3000	0.1397	0.14	0.258	0.199	0.033	0.213	0.123
0.3443	0.7	3500	0.1333	0.139	0.044	0.092	0.032	0.027	0.029
0.3169	0.8	4000	0.1297	0.129	0.027	0.078	0.029	0.011	0.02
0.3276	0.9	4500	0.1264	0.124	0.086	0.105	0.028	0.058	0.043
0.317	1.0	5000	0.1255	0.124	0.171	0.147	0.028	0.164	0.096

Framework versions

• Transformers 4.42.4
• Pytorch 2.2.0
• Datasets 2.20.0
• Tokenizers 0.19.1