whisper-small-telugu

This model is a fine-tuned version of openai/whisper-small on the google/fleurs dataset. It achieves the following results on the evaluation set (google/flerus telugu test set):

• Loss: 0.3622
• Wer: 39.6774

openai/whisper-small has the following zero shot performance on google/fleurs test set:

• Wer: 117.91

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2623	1.55	500	0.2733	65.9750
0.0859	3.1	1000	0.2045	39.7652
0.0538	4.64	1500	0.2220	42.3811
0.0265	6.19	2000	0.2526	42.3626
0.0179	7.74	2500	0.2754	42.1685
0.008	9.29	3000	0.2966	41.2257
0.0061	10.83	3500	0.2950	40.6202
0.0034	12.38	4000	0.3049	40.3198
0.004	13.93	4500	0.3106	40.5879
0.0018	15.48	5000	0.3199	40.1812
0.0016	17.03	5500	0.3346	39.8345
0.0006	18.57	6000	0.3337	40.2274
0.0003	20.12	6500	0.3396	40.2597
0.0005	21.67	7000	0.3465	40.1072
0.0002	23.22	7500	0.3485	39.7282
0.0002	24.77	8000	0.3519	39.7837
0.0001	26.32	8500	0.3567	39.7560
0.0001	27.86	9000	0.3614	39.8068
0.0	29.41	9500	0.3609	39.4925
0.0	30.96	10000	0.3622	39.6774

Framework versions

• Transformers 4.24.0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1
• Tokenizers 0.13.2