whisper-large-v2-mn-13

This model is a fine-tuned version of openai/whisper-large-v2 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1689
• Wer: 20.0240
• Cer: 6.6010

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: 8
• eval_batch_size: 4
• 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: 25000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
0.3921	0.09	1000	15.7845	0.4101	46.9030
0.3115	0.17	2000	14.2911	0.3353	41.8451
0.2659	0.26	3000	11.8131	0.2800	34.6406
0.2477	0.35	4000	10.6659	0.2578	32.0024
0.2274	0.43	5000	10.0460	0.2463	30.3419
0.2059	0.52	6000	9.9264	0.2305	28.5558
0.2092	0.61	7000	9.4277	0.2196	27.8785
0.1956	0.69	8000	9.2745	0.2093	26.8353
0.195	0.78	9000	8.9485	0.2042	26.6168
0.195	0.87	10000	8.5324	0.2001	25.6718
0.1795	0.95	11000	8.1786	0.1936	24.1698
0.1575	1.04	12000	7.8653	0.1915	23.8912
0.1358	1.13	13000	7.6749	0.1918	23.3778
0.1509	1.21	14000	7.7221	0.1852	23.1811
0.1474	1.3	15000	7.3246	0.1764	22.4984
0.1461	1.39	16000	7.3187	0.1793	22.4110
0.134	1.47	17000	7.1123	0.1737	21.9412
0.1289	1.56	18000	7.4593	0.1727	22.0614
0.1287	1.65	19000	7.0230	0.1701	21.4223
0.1196	1.73	20000	6.9447	0.1666	21.2475
0.1275	1.82	21000	6.7956	0.1653	20.8106
0.1329	1.91	22000	6.7729	0.1622	20.3354
0.1294	1.99	23000	6.6448	0.1606	20.2207
0.1043	2.08	24000	6.6010	0.1689	20.0240
0.079	2.17	25000	6.6246	0.1687	20.1005

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.8.1.dev0
• Tokenizers 0.13.2