deepdml/whisper-small-ar-mix-norm

Whisper Small ar

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

• Loss: 0.3340
• Wer: 27.1515
• Cer: 7.9929

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: 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
0.5511	0.0556	1000	0.4263	38.7837	12.1682
0.2273	0.1111	2000	0.3858	34.5513	10.7969
0.1023	0.1667	3000	0.3663	33.5690	10.3863
0.0545	0.2222	4000	0.3567	31.5786	9.2661
0.043	0.2778	5000	0.3421	31.7236	9.3731
0.0254	0.3333	6000	0.3316	30.0600	9.0426
0.0219	0.3889	7000	0.3269	29.6451	8.7922
0.0177	0.4444	8000	0.3258	29.2705	8.7774
0.0209	0.5	9000	0.3157	28.5177	8.5056
0.0212	0.5556	10000	0.3105	28.9345	8.4034
0.0093	0.6111	11000	0.3111	27.8052	8.1165
0.012	0.6667	12000	0.3158	27.7042	8.2345
0.0124	0.7222	13000	0.3119	27.0304	7.9191
0.005	1.0393	14000	0.3392	27.5739	8.0862
0.008	1.0949	15000	0.3334	27.3590	7.9829
0.0049	1.1504	16000	0.3451	27.2911	8.0076
0.0032	1.206	17000	0.3468	27.0873	7.9693
0.0031	1.2616	18000	0.3340	27.1515	7.9929

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