b-brave/asr2_medium_v0.9

Whisper Medium

This model is a fine-tuned version of openai/whisper-medium on the b-brave-clean dataset. It achieves the following results on the evaluation set:

• Loss: 0.6398
• Wer: 41.8338
• Cer: 30.4702
• Lr: 0.0000

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: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.3
• num_epochs: 12
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Lr
3.6633	1.0	168	1.8570	149.4269	97.5309	0.0001
1.2316	2.0	336	0.9484	66.4756	45.5477	0.0002
0.8773	3.0	504	0.8240	117.3352	102.2590	0.0002
0.5747	4.0	672	0.7314	76.3610	56.0809	0.0003
0.3652	5.0	840	0.6545	104.1547	90.4124	0.0002
0.2632	6.0	1008	0.6305	50.5731	34.6467	0.0002
0.1568	7.0	1176	0.5926	90.1146	80.4833	0.0002
0.104	8.0	1344	0.6388	47.5645	33.4121	0.0001
0.0356	9.0	1512	0.6073	42.8367	46.9398	0.0001
0.0189	10.0	1680	0.6361	42.5501	47.1500	0.0001
0.0089	11.0	1848	0.6385	41.9771	30.4965	0.0000
0.0108	11.9313	2004	0.6398	41.8338	30.4702	0.0000

Framework versions

• PEFT 0.14.0
• Transformers 4.48.3
• Pytorch 2.2.0
• Datasets 3.2.0
• Tokenizers 0.21.0