Kushtrim/whisper-large-v3-turbo-sq

Whisper Turbo SQ

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Common Voice 19.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0880
• Wer: 7.0282

Model description

You can read more about the model in the openai/whisper-large-v3-turbo model card.

Performance and Limitations

The Whisper Large V3 Turbo SQ model demonstrates improved performance compared to the pretrained version, with a WER of 7.02. However, the results are not yet optimal. The main challenge is the lack of sufficient and high-quality data for the Albanian language. This model serves as an example to highlight that increased community participation and voice donations can significantly enhance performance. To achieve top-tier results, a larger and more diverse dataset is essential. Contributions from the community are crucial for improving the model's accuracy and efficiency. You can contribute to this effort by visiting the Mozilla Common Voice website.

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 4
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3481	0.4237	250	0.3336	28.2712
0.2877	0.8475	500	0.2711	26.1243
0.1694	1.2712	750	0.2183	19.8079
0.1405	1.6949	1000	0.1722	15.8757
0.0623	2.1186	1250	0.1430	13.7514
0.0594	2.5424	1500	0.1238	12.4633
0.0379	2.9661	1750	0.1054	9.6949
0.017	3.3898	2000	0.0968	8.3051
0.0133	3.8136	2250	0.0880	7.0282

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.0+cu121
• Datasets 3.0.1
• Tokenizers 0.20.0