Whisper-large-v3-turbo-pa

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

• Loss: 0.043609
• Wer: 14.618034

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: 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_steps: 500
• training_steps: 18000
• mixed_precision_training: Native AMP

Initial Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0974	1.2804	1000	0.1187	29.9568
0.0669	2.5608	2000	0.0974	25.7646
0.045	3.8412	3000	0.0938	24.6229
0.0207	5.1216	4000	0.1090	24.4019
0.0159	6.4020	5000	0.1179	23.8895
0.0094	7.6825	6000	0.1273	23.3044
0.005	8.9629	7000	0.1387	23.0443
0.0014	10.2433	8000	0.1502	22.5476
0.0007	11.5237	9000	0.1589	22.2849
0.0002	12.8041	10000	0.1607	22.1003

Re-Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.070200	-	1000	0.056950	17.876652
0.063900	-	2000	0.057231	17.571240
0.055600	-	3000	0.055906	17.123214
0.051300	-	4000	0.053482	17.294083
0.050100	-	5000	0.051127	16.434357
0.045300	-	6000	0.048763	15.636520
0.046700	-	7000	0.046148	15.137368
0.037200	-	8000	0.043609	14.618034

Framework versions

• Transformers 4.46.3
• Pytorch 2.1.0+cu118
• Datasets 3.1.0
• Tokenizers 0.20.3