whisper-small-noisy-hindi

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: 1.6842
• Wer: 76.4655

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: 3000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.5492	0.61	50	2.2588	122.9379
1.6655	1.22	100	1.6099	92.4434
1.4463	1.83	150	1.4333	89.4086
1.2723	2.44	200	1.3012	90.1954
1.1433	3.05	250	1.1803	84.7052
0.9489	3.66	300	1.0427	82.0768
0.8082	4.27	350	0.9796	81.3246
0.7696	4.88	400	0.9372	78.6962
0.631	5.49	450	0.9229	78.2898
0.5784	6.1	500	0.9306	76.3790
0.5112	6.71	550	0.9158	76.3963
0.3738	7.32	600	0.9585	75.7652
0.3762	7.93	650	0.9530	75.5490
0.2647	8.54	700	1.0094	76.1975
0.2282	9.15	750	1.0548	76.8632
0.1865	9.76	800	1.0789	76.4482
0.1172	10.37	850	1.1491	78.1688
0.1181	10.98	900	1.1769	76.5520
0.0778	11.59	950	1.2255	77.5117
0.0579	12.2	1000	1.3021	76.0246
0.0515	12.8	1050	1.3064	76.8546
0.0324	13.41	1100	1.3766	77.1140
0.0317	14.02	1150	1.4044	78.2206
0.0227	14.63	1200	1.4420	77.9353
0.0174	15.24	1250	1.4780	76.4482
0.0164	15.85	1300	1.5044	76.2494
0.0121	16.46	1350	1.5338	76.7595
0.0128	17.07	1400	1.5588	76.9410
0.0108	17.68	1450	1.5688	76.4482
0.0085	18.29	1500	1.6060	76.6903
0.0082	18.9	1550	1.6368	76.5606
0.0065	19.51	1600	1.6483	76.5520
0.0062	20.12	1650	1.6842	76.4655

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 1.12.1
• Datasets 2.16.1
• Tokenizers 0.15.0