learning_rate: 1e-05
train_batch_size: 32
eval_batch_size: 32
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
0.1548	0.0556	1000	0.2268	46.2895	9.6201
0.1066	0.1111	2000	0.2074	42.7291	8.5157
0.0857	0.1667	3000	0.1900	39.2279	7.6586
0.0702	0.2222	4000	0.1748	38.1497	7.1386
0.0749	0.2778	5000	0.1721	37.2119	7.1248
0.0538	0.3333	6000	0.1791	36.8448	6.9975
0.0612	0.3889	7000	0.1684	36.0165	6.7150
0.0435	0.4444	8000	0.1724	35.1999	6.6523
0.0601	0.5	9000	0.1605	35.2295	6.5536
0.0553	0.5556	10000	0.1543	34.3342	6.3516
0.0431	0.6111	11000	0.1535	33.5807	6.1092
0.0375	0.6667	12000	0.1572	33.5575	6.2222
0.0566	0.7222	13000	0.1582	33.5549	6.1894
0.034	0.7778	14000	0.1568	33.0499	5.9889
0.0464	0.8333	15000	0.1578	32.8799	5.9867
0.0527	0.8889	16000	0.1580	32.8477	6.0074
0.0378	0.9444	17000	0.1632	32.6918	5.9512
0.031	1.0	18000	0.1594	32.5592	5.8817

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:

bibtex

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