arielcerdap/whisper-largev3-fluencybank

Whisper fine-tuned on FluencyBank — openai/whisper-large-v3

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

• Loss: 1.8275
• Wer: 9.7561
• Cer: 5.7991

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: 8e-06
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• training_steps: 2500
• label_smoothing_factor: 0.1

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.4541	11.6279	250	1.6985	11.6678	7.8383
1.4269	23.2558	500	1.7426	9.6902	5.6261
1.4251	34.8837	750	1.7460	9.7781	5.7536
1.42	46.5116	1000	1.7811	9.6243	5.5760
1.4199	58.1395	1250	1.7827	9.8220	5.7991
1.418	69.7674	1500	1.8085	9.5803	5.6580
1.4176	81.3953	1750	1.8191	9.8000	5.7354
1.4173	93.0233	2000	1.8252	9.7341	5.7445
1.4173	104.6512	2250	1.8272	9.8220	5.8446
1.4172	116.2791	2500	1.8275	9.7561	5.7991

Framework versions

• Transformers 4.45.2
• Pytorch 2.10.0+cu128
• Datasets 4.0.0
• Tokenizers 0.20.3