whisper-base-fa - Sadegh Karimi

This model is a fine-tuned version of SadeghK/whisper-base on the Common Voice 20.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0813
• Wer: 10.3712

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Use convert-to-ggml.ipynb to convert to ggml

To run faster with whisper.cpp, use convert-to-ggml.ipynb to convert model. Model is already converted and saved as "ggml-base-fa.bin"

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: 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_steps: 500
• training_steps: 50000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1234	0.0493	1000	0.1698	21.8312
0.1303	0.0986	2000	0.1663	22.9153
0.1241	0.1479	3000	0.1623	20.8843
0.1223	0.1972	4000	0.1616	20.7470
0.1281	0.2465	5000	0.1522	19.3606
0.1111	0.2958	6000	0.1483	20.0901
0.1097	0.3451	7000	0.1452	19.0445
0.1439	0.3944	8000	0.1367	18.0251
0.1053	0.4437	9000	0.1347	17.5902
0.1248	0.4930	10000	0.1281	16.9486
0.1081	0.5423	11000	0.1252	15.9200
0.1062	0.5916	12000	0.1222	15.8167
0.1139	0.6409	13000	0.1181	15.6038
0.1011	0.6902	14000	0.1145	15.0918
0.098	0.7395	15000	0.1141	15.0194
0.1176	0.7888	16000	0.1091	14.1048
0.0933	0.8381	17000	0.1067	13.9028
0.0981	0.8874	18000	0.1042	13.6391
0.0909	0.9367	19000	0.1012	13.2119
0.0714	0.9860	20000	0.1001	13.1826
0.0491	1.0353	21000	0.0985	12.9251
0.059	1.0846	22000	0.0966	12.6799
0.0492	1.1339	23000	0.0959	12.4501
0.0625	1.1832	24000	0.0943	12.5241
0.0429	1.2325	25000	0.0946	12.4424
0.0403	1.2818	26000	0.0931	12.1370
0.0474	1.3311	27000	0.0921	11.7330
0.0484	1.3804	28000	0.0910	11.5710
0.0585	1.4297	29000	0.0896	11.7067
0.0431	1.4790	30000	0.0890	11.3875
0.045	1.5283	31000	0.0875	11.2842
0.0494	1.5776	32000	0.0862	11.5433
0.0448	1.6269	33000	0.0854	11.0282
0.0508	1.6762	34000	0.0849	11.0498
0.0432	1.7255	35000	0.0837	10.7583
0.0356	1.7748	36000	0.0826	10.8339
0.0353	1.8241	37000	0.0819	10.5300
0.043	1.8734	38000	0.0815	10.4838
0.0434	1.9227	39000	0.0812	10.5038
0.0382	1.9720	40000	0.0809	10.4684
0.0342	2.0213	41000	0.0833	10.4853
0.0249	2.0706	42000	0.0841	10.7783
0.0237	2.1199	43000	0.0835	10.5100
0.0282	2.1692	44000	0.0835	10.5563
0.0277	2.2185	45000	0.0830	10.7151
0.0328	2.2678	46000	0.0824	10.3959
0.0268	2.3171	47000	0.0822	10.4560
0.0395	2.3664	48000	0.0817	10.3311
0.0298	2.4157	49000	0.0815	10.4128
0.029	2.4650	50000	0.0813	10.3712

Framework versions

• Transformers 4.48.2
• Pytorch 2.1.0+cu118
• Datasets 3.2.0
• Tokenizers 0.21.0