arielcerdap/largev3-turbo-stutter

Large V3 Turbo Stutter - Ariel Cerda

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Libristutter 4.7k dataset. It achieves the following results on the evaluation set:

• Loss: 0.5114
• Wer: 24.2314

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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0584	3.7453	1000	0.3100	20.0250
0.0117	7.4906	2000	0.3970	21.1392
0.002	11.2360	3000	0.4427	20.3825
0.0003	14.9813	4000	0.4927	23.8501
0.0002	18.7266	5000	0.5114	24.2314

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.19.1