whisper-NST2-unfreeze-constanti-low-lr

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

• Loss: 0.3562
• Wer: 8.5519

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: 96
• 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
0.1901	0.05	1000	0.3069	14.8233
0.1323	0.1	2000	0.2687	11.2885
0.1137	0.15	3000	0.2620	10.8324
0.1022	0.2	4000	0.2976	9.0080
0.0937	0.25	5000	0.2584	9.5781
0.0875	0.3	6000	0.2704	20.2965
0.0592	1.05	7000	0.2751	9.0080
0.0488	1.1	8000	0.2778	8.6659
0.0475	1.15	9000	0.2792	9.4641
0.0439	1.2	10000	0.2880	8.3238
0.0425	1.25	11000	0.2954	8.5519
0.0416	1.3	12000	0.2896	20.2965
0.0289	2.05	13000	0.2990	7.9818
0.0229	2.1	14000	0.3027	7.4116
0.0248	2.15	15000	0.2968	8.6659
0.0225	2.2	16000	0.3100	8.5519
0.0222	2.25	17000	0.3132	9.3501
0.0219	2.3	18000	0.3230	7.6397
0.0162	3.04	19000	0.3380	9.8062
0.0132	3.09	20000	0.3562	8.5519

Framework versions

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