whisper-small-telugu-large-data

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

• Loss: 0.3310
• Wer: 38.8460

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: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.128	2.27	500	0.2015	45.1692
0.0462	4.55	1000	0.1877	41.1050
0.0184	6.82	1500	0.2241	40.5153
0.0045	9.09	2000	0.2590	39.7260
0.0019	11.36	2500	0.2824	39.0819
0.0006	13.64	3000	0.3002	38.9096
0.0002	15.91	3500	0.3141	38.5920
0.0001	18.18	4000	0.3232	38.7463
0.0001	20.45	4500	0.3289	38.8370
0.0001	22.73	5000	0.3310	38.8460

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.0+cu117
• Datasets 2.7.1
• Tokenizers 0.13.2