Modelpublic

Khalsuu/filipino-wav2vec2-l-xls-r-300m-official

sourceHugging Faceapache-2.0updated 4y agoView on Hugging Face

2likes1.7mdownloads

Model Card

filipino-wav2vec2-l-xls-r-300m-official

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the filipino_voice dataset. It achieves the following results on the evaluation set:

—Loss: 0.4672
—Wer: 0.2922

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: 0.0003
—trainbatchsize: 8
—evalbatchsize: 8
—seed: 42
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 16
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: linear
—lrschedulerwarmup_steps: 500
—num_epochs: 30
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.3671	2.09	400	0.5584	0.5987
0.48	4.19	800	0.4244	0.4195
0.2796	6.28	1200	0.3742	0.3765
0.1916	8.38	1600	0.4291	0.3667
0.1463	10.47	2000	0.3745	0.3415
0.1165	12.57	2400	0.4472	0.3407
0.0955	14.66	2800	0.4269	0.3290
0.0823	16.75	3200	0.4608	0.3475
0.0709	18.85	3600	0.4706	0.3281
0.0603	20.94	4000	0.4380	0.3183
0.0527	23.04	4400	0.4473	0.3067
0.0449	25.13	4800	0.4550	0.3029
0.041	27.23	5200	0.4671	0.3020
0.0358	29.32	5600	0.4672	0.2922

Framework versions

—Transformers 4.11.3
—Pytorch 1.10.0+cu113
—Datasets 1.18.3
—Tokenizers 0.10.3