Modelpublic

Miniijune/wav2vec2-xls-r-300m-Korean-children-pronunciation

sourceHugging Faceapache-2.0updated 9mo agoView on Hugging Face

0likes7downloads

Model Card

wav2vec2-xls-r-300m-Korean-children-pronunciation

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the None dataset. It achieves the following results on the evaluation set:

The goal of my model is to transfer the voice to the Korean pronunciation.

The model is trained by using Korean-Childern Pronuncitation in AI-Hub

I converted Korean-Sentence to Korean-pronunciation using G2PK python library

More information needed

Train - 250,000

Validation - 5,000

Test - 5,000

It took about 32h to learn

The following hyperparameters were used during training:

—learning_rate: 1e-4
—trainbatchsize: 4
—evalbatchsize: 1
—seed: 42
—gradientaccumulationsteps: 8
—totaltrainbatch_size: 32
—optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
—lrschedulertype: linear
—lrschedulerwarmup_steps: 2000
—num_epochs: 5
—mixedprecisiontraining: Native AMP

Training Loss	Epoch	Step	Validation Loss	Cer
2.7335	0.2568	2000	2.9172	0.4084
1.401	0.5136	4000	1.9700	0.2520
1.1359	0.7703	6000	1.5784	0.2201
0.9546	1.0271	8000	1.3394	0.1981
0.9167	1.2839	10000	1.2866	0.1901
0.8224	1.5406	12000	1.2323	0.1795
0.7777	1.7974	14000	1.1941	0.1695
0.7436	2.0542	16000	1.2106	0.1676
0.7119	2.3110	18000	1.2312	0.1608
0.6783	2.5677	20000	1.1446	0.1566
0.6452	2.8245	22000	1.1421	0.1513
0.6208	3.0813	24000	1.1602	0.1460
0.5758	3.3380	26000	1.1391	0.1435
0.5673	3.5948	28000	1.0839	0.1395
0.5353	3.8516	30000	1.0952	0.1364
0.5235	4.1084	32000	1.0479	0.1343
0.5056	4.3651	34000	1.0552	0.1312
0.4978	4.6219	36000	1.0342	0.1292
0.4895	4.8787	38000	1.0513	0.1287