whisper-large-v3-turbo-common_voice_16_1-zh-TW-2

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the JacobLinCool/mozilla-foundation-common_voice_16_1-zh-TW-preprocessed dataset. It achieves the following results on the evaluation set:

• Loss: 0.2346
• Wer: 38.5450
• Cer: 10.8963

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: 0.0005
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
No log	0	0	2.7503	76.5675	20.3917
0.9352	0.9987	377	0.2472	47.9301	13.6656
0.73	1.9980	754	0.2502	47.0056	13.5652
0.4985	2.9974	1131	0.2559	46.2018	13.7057
0.1928	3.9993	1509	0.2595	45.9606	13.0906
0.2539	4.9987	1886	0.2522	44.7950	13.1459
0.0607	5.9980	2263	0.2422	44.7548	12.5006
0.0826	6.9974	2640	0.2488	43.8907	12.4906
0.0151	7.9993	3018	0.2403	40.2331	11.4537
0.0056	8.9987	3395	0.2390	39.8312	11.5290
0.0056	9.9927	3770	0.2346	38.5450	10.8963

Framework versions

• PEFT 0.13.2
• Transformers 4.46.0
• Pytorch 2.4.1+cu124
• Datasets 3.0.2
• Tokenizers 0.20.1