valerdon/whisper-large-v2-050324

whisper-large-v2

This model is a fine-tuned version of openai/whisper-large-v2 on the FS_phone_calls dataset. It achieves the following results on the evaluation set:

• Loss: 3.0867
• Wer: 98.9696

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: 5
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.8251	3.33	100	1.7838	93.7146
1.0083	6.67	200	2.0824	92.7872
0.4413	10.0	300	2.2075	98.6090
0.131	13.33	400	2.5746	103.8125
0.0484	16.67	500	2.6733	107.2128
0.031	20.0	600	2.6507	99.4333
0.022	23.33	700	2.6682	103.2457
0.013	26.67	800	2.7772	105.3581
0.0135	30.0	900	2.7849	100.3091
0.0097	33.33	1000	2.7935	96.0845
0.005	36.67	1100	2.9164	94.8480
0.0039	40.0	1200	2.8849	100.8758
0.0031	43.33	1300	2.9600	100.4637
0.0013	46.67	1400	2.9947	104.1731
0.001	50.0	1500	3.0367	100.1030
0.001	53.33	1600	3.0172	95.2602
0.0008	56.67	1700	3.0539	100.2061
0.0007	60.0	1800	3.0730	100.4122
0.0006	63.33	1900	3.0813	99.1757
0.0006	66.67	2000	3.0867	98.9696

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2