Modelpublic

kiranpantha/whisper-medium-nepali

sourceHugging Faceapache-2.0updated 2y agoView on Hugging Face

0likes20downloads

Model Card

Whisper Medium Nepali - Kiran Pantha

This model is a fine-tuned version of openai/whisper-medium on the OpenSLR54 dataset. It achieves the following results on the evaluation set:

—Loss: 0.1608
—Wer: 21.9907
—Cer: 5.3068

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
—trainbatchsize: 16
—evalbatchsize: 8
—seed: 42
—optimizer: Use adamwtorch with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
—lrschedulertype: linear
—lrschedulerwarmup_steps: 500
—training_steps: 5000
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
0.2877	0.3597	300	13.1055	0.2588	48.6574
0.1865	0.7194	600	10.0741	0.2052	39.9537
0.1155	1.0791	900	8.1217	0.1633	31.8981
0.0992	1.4388	1200	7.6923	0.1577	30.6250
0.094	1.7986	1500	7.1969	0.1418	28.3796
0.0454	2.1583	1800	6.7308	0.1472	26.6898
0.0333	2.5180	2100	6.6353	0.1512	27.0602
0.0446	2.8777	2400	6.25	0.1409	25.5556
0.0204	3.2374	2700	6.5399	0.1513	25.8796
0.016	3.5971	3000	5.9674	0.1560	24.8380
0.0166	3.9568	3300	6.0592	0.1573	25.0926
0.0052	4.3165	3600	5.7802	0.1566	23.3796
0.0068	4.6763	3900	5.5233	0.1544	22.7083
0.0013	5.0360	4200	5.4756	0.1568	22.7546
0.001	5.3957	4500	0.1606	22.0139	5.3178
0.0017	5.7554	4800	0.1608	21.9907	5.3068

Framework versions

—Transformers 4.46.3
—Pytorch 2.5.1+cxx11.abi
—Datasets 3.2.0
—Tokenizers 0.20.3