whisper-medium-NST-uf-linlr

This model is a fine-tuned version of openai/whisper-medium on the NBAILAB/NST - NO-CLOSE dataset. It achieves the following results on the evaluation set:

• Loss: 0.3007
• Wer: 9.1220

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: 72
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• training_steps: 20000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2046	0.05	1000	0.3426	15.2794
0.148	0.1	2000	0.3284	10.8324
0.121	0.15	3000	0.3092	12.8848
0.1089	0.2	4000	0.2808	10.4903
0.0976	0.25	5000	0.2617	9.9202
0.0901	0.3	6000	0.2604	21.8928
0.0834	0.35	7000	0.2877	9.3501
0.0825	0.4	8000	0.2794	9.3501
0.0553	1.05	9000	0.2845	9.5781
0.0472	1.1	10000	0.2814	24.1733
0.0409	1.15	11000	0.3084	8.0958
0.041	1.2	12000	0.2865	9.2360
0.0353	1.25	13000	0.2828	6.4994
0.0348	1.3	14000	0.2708	7.5257
0.0349	1.35	15000	0.2842	23.0331
0.0361	1.4	16000	0.2769	10.1482
0.0249	2.04	17000	0.2935	8.8940
0.0204	2.09	18000	0.2874	12.4287
0.0175	2.14	19000	0.2882	12.9989
0.0197	2.19	20000	0.3007	9.1220

Framework versions

• Transformers 4.25.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.6.1
• Tokenizers 0.13.1