whisper-small-Denoiser-enhanced-weight-05-05-hindi-10dB

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

• Loss: 0.5849
• Wer: 34.2815

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: 64
• eval_batch_size: 32
• 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: 1650
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.6592	0.61	50	1.3742	83.7973
0.8218	1.22	100	0.7806	57.9198
0.642	1.83	150	0.6349	58.6201
0.504	2.44	200	0.5347	54.5651
0.406	3.05	250	0.4393	43.1264
0.2617	3.66	300	0.3305	40.6277
0.1785	4.27	350	0.3107	39.2703
0.1634	4.88	400	0.2939	37.7313
0.1042	5.49	450	0.3005	37.7572
0.0913	6.1	500	0.3094	36.2528
0.0612	6.71	550	0.3192	36.3566
0.033	7.32	600	0.3379	35.8378
0.0332	7.93	650	0.3420	34.9818
0.021	8.54	700	0.3562	35.1288
0.0143	9.15	750	0.3713	35.3017
0.0109	9.76	800	0.3667	34.7657
0.0073	10.37	850	0.3885	35.6562
0.0075	10.98	900	0.3953	34.4631
0.0042	11.59	950	0.4094	34.5582
0.0036	12.2	1000	0.4179	34.1605
0.0028	12.8	1050	0.4307	34.3247
0.0028	13.41	1100	0.4399	34.2383
0.0014	14.02	1150	0.4490	34.1691
0.0015	14.63	1200	0.4682	34.6187
0.0005	15.24	1250	0.4833	34.3680
0.0008	15.85	1300	0.4916	34.0913
0.0005	16.46	1350	0.5065	33.9270
0.0004	17.07	1400	0.5176	34.1345
0.0002	17.68	1450	0.5429	34.2988
0.0001	18.29	1500	0.5548	33.8233
0.0	18.9	1550	0.5669	34.3075
0.0	19.51	1600	0.5814	34.4631
0.0	20.12	1650	0.5849	34.2815

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 1.12.0+cu113
• Datasets 2.16.1
• Tokenizers 0.15.0