deepdml/whisper-large-v3-turbo-ta-mix-norm

Whisper Turbo ta

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1246
• Wer: 28.7424
• Cer: 4.7569

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: 16
• 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_ratio: 0.04
• training_steps: 18000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.2207	0.0556	1000	0.2388	48.5270	9.7063
0.1306	0.1111	2000	0.2122	43.2340	8.2086
0.1075	0.1667	3000	0.2010	41.0442	7.7213
0.1279	0.2222	4000	0.1906	38.8311	7.2164
0.1084	0.2778	5000	0.1820	37.9591	6.9916
0.0707	0.3333	6000	0.1701	35.2656	6.3098
0.0676	0.3889	7000	0.1611	34.9628	6.1673
0.0753	0.4444	8000	0.1571	34.1384	5.9429
0.0714	0.5	9000	0.1573	33.8318	6.0002
0.0832	0.5556	10000	0.1572	33.6309	5.8641
0.0663	0.6111	11000	0.1488	32.9469	5.6721
0.0488	0.6667	12000	0.1475	32.0426	5.5184
0.0462	0.7222	13000	0.1466	31.6781	5.3493
0.0362	0.7778	14000	0.1441	31.1512	5.2179
0.0528	0.8333	15000	0.1302	29.9777	4.9914
0.0507	0.8889	16000	0.1304	29.8051	4.9342
0.0451	0.9444	17000	0.1275	29.5191	4.8730
0.0425	1.0	18000	0.1246	28.7424	4.7569

Framework versions

• Transformers 4.48.0.dev0
• Pytorch 2.5.1+cu121
• Datasets 3.6.0
• Tokenizers 0.21.0

Citation

Please cite the model using the following BibTeX entry:

@misc{deepdml/whisper-large-v3-turbo-ta-mix-norm,
      title={Fine-tuned Whisper turbo ASR model for speech recognition in Tamil},
      author={Jimenez, David},
      howpublished={\url{https://huggingface.co/deepdml/whisper-large-v3-turbo-ta-mix-norm}},
      year={2026}
    }