michaelszhu/whisper-small-fintuned-radio-ASR

BANG - v2 (EN)

This model is a fine-tuned version of openai/whisper-large-v3 on the Radio-Modified Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2650
• Wer: 20.5610

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: 16
• eval_batch_size: 8
• 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: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.7709	0.25	1000	0.6383	35.6607
0.4424	1.2443	2000	0.4248	26.8037
0.2823	2.2385	3000	0.3117	22.4425
0.2429	3.2328	4000	0.2650	20.5610

Framework versions

• Transformers 4.41.1
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1