whisper-large-v3-turbo-finetune-20-sw

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the common_voice_17_0 dataset. If it useful to you. pls give me a like. thanks It achieves the following results on the evaluation set:

• Best WER(Validation): 17.62
• Best CER(Validation): 5.41
• Target Language: Swahili

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: 4
• eval_batch_size: 2
• gradient_accumulation_steps: 8
• bf16: true
• warmup_ratio: 0.1
• 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: 20
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	WER	CER
1000	0.332600	0.454433	27.630518	8.602336
2000	0.225900	0.403464	25.752715	8.555724
3000	0.168600	0.376518	22.879734	7.451020
4000	0.180700	0.364340	23.105783	8.354530
5000	0.116200	0.361309	21.115426	7.246847
6000	0.068100	0.383298	21.475604	7.591304
7000	0.073200	0.369184	19.368938	6.654584
8000	0.045300	0.395950	18.973118	5.780262
9000	0.035800	0.399188	18.650459	5.684506
10000	0.035800	0.386173	18.434727	5.821514
11000	0.022500	0.424523	18.192733	5.648466
12000	0.011900	0.450201	18.242445	5.576984
13000	0.017900	0.422032	18.033279	5.705504
14000	0.009200	0.441910	17.773463	5.562687
15000	0.006800	0.472621	17.625265	5.412574
16000	0.004800	0.454455	18.644831	6.194713

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.3.1
• Tokenizers 0.21.0