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

Whisper Turbo af

This model is a fine-tuned version of deepdml/whisper-large-v3-turbo on multiple datasets. It achieves the following results on the evaluation set:

• Loss: 0.6490
• Wer: 20.7864
• Cer: 7.3989

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.04
• training_steps: 4100

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.3922	0.0244	100	0.5760	26.3988	8.2784
0.3225	0.0488	200	0.5829	26.6932	10.3802
0.2975	0.0732	300	0.5755	27.1263	10.5238
0.2038	0.0976	400	0.5738	27.2475	11.5557
0.1537	0.1220	500	0.5849	26.0870	10.5151
0.1461	0.1463	600	0.6038	25.9137	9.1080
0.1342	0.1707	700	0.5824	25.1516	9.0288
0.1091	0.1951	800	0.5955	22.9517	8.1934
0.1215	0.2195	900	0.5839	23.6619	7.7654
0.0973	0.2439	1000	0.5974	22.9517	8.0204
0.063	0.2683	1100	0.6047	22.3627	7.4341
0.0678	0.2927	1200	0.6247	22.4493	7.8005
0.0598	0.3171	1300	0.6182	22.1549	7.9266
0.0553	0.3415	1400	0.6137	22.4147	8.0204
0.0514	0.3659	1500	0.6539	23.0383	7.8211
0.0409	0.3902	1600	0.6350	22.9170	7.5807
0.0496	0.4146	1700	0.6281	21.9297	7.5367
0.0607	0.4390	1800	0.6213	23.2808	9.0464
0.0321	0.4634	1900	0.6303	21.6352	7.5602
0.0301	0.4878	2000	0.6395	23.6099	8.4982
0.0454	0.5122	2100	0.6431	23.3674	7.9471
0.0236	0.5366	2200	0.6437	21.5659	7.3374
0.0308	0.5610	2300	0.6215	22.9517	7.3051
0.019	0.5854	2400	0.6628	23.1942	7.8885
0.0255	0.6098	2500	0.6488	24.3028	9.2575
0.0307	0.6341	2600	0.6325	24.7358	9.1373
0.0305	0.6585	2700	0.6382	23.1942	8.5891
0.021	0.6829	2800	0.6510	22.9170	8.4836
0.0177	0.7073	2900	0.6525	22.1375	7.3081
0.0172	0.7317	3000	0.6500	22.3108	7.5396
0.0172	0.7561	3100	0.6569	21.6352	7.4957
0.0133	0.7805	3200	0.6506	21.9297	8.6624
0.0155	0.8049	3300	0.6425	20.9943	7.2143
0.0211	0.8293	3400	0.6482	20.9770	7.1644
0.0189	0.8537	3500	0.6588	20.6998	7.1058
0.021	0.8780	3600	0.6480	20.5786	7.2436
0.0124	0.9024	3700	0.6547	20.9423	7.1322
0.0125	0.9268	3800	0.6607	20.9943	6.9827
0.0072	0.9512	3900	0.6571	21.0982	6.9827
0.0181	0.9756	4000	0.6500	20.9423	7.4517
0.0091	1.0	4100	0.6490	20.7864	7.3989

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1

Citation

Please cite the model using the following BibTeX entry:

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