./whisper-large-cit-synth-do015-wd0-lr5e-06-1000

This model is a fine-tuned version of openai/whisper-large-v3 on the SF 1000 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4526
• Wer: 20.3899

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: 5e-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.7187	0.8889	50	0.4062	24.2105
0.4122	1.7778	100	0.3523	22.3782
0.2917	2.6667	150	0.3494	23.5867
0.2242	3.5556	200	0.3618	23.0019
0.1529	4.4444	250	0.3770	22.3392
0.1322	5.3333	300	0.3906	21.2476
0.0987	6.2222	350	0.4133	20.9747
0.0798	7.1111	400	0.4302	23.8986
0.0613	8.0	450	0.4438	20.5848
0.0545	8.8889	500	0.4526	20.3899

Framework versions

• Transformers 4.42.3
• Pytorch 1.13.1+cu117
• Datasets 2.20.0
• Tokenizers 0.19.1