Modelpublic

RuiqianLi/Malaya-speech_fine-tune_MrBrown_15_Jun

sourceHugging Faceupdated 4y agoView on Hugging Face

0likes19downloads

Model Card

Malaya-speechfine-tuneMrBrown15Jun

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

—Loss: 0.4822
—Wer: 0.2449

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: 0.0001
—trainbatchsize: 2
—evalbatchsize: 8
—seed: 42
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 4
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: linear
—lrschedulerwarmup_steps: 1000
—num_epochs: 100
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.1607	5.26	200	0.3983	0.2381
0.5184	10.52	400	0.3256	0.2245
0.2993	15.78	600	0.3437	0.2426
0.2485	21.05	800	0.4547	0.2585
0.1917	26.31	1000	0.4598	0.2517
0.1586	31.57	1200	0.4050	0.2290
0.1486	36.83	1400	0.4186	0.2653
0.1307	42.1	1600	0.4284	0.2857
0.0895	47.36	1800	0.5158	0.2562
0.0526	52.62	2000	0.4525	0.2449
0.0553	57.88	2200	0.4364	0.2336
0.037	63.16	2400	0.3873	0.2449
0.0439	68.42	2600	0.3914	0.2404
0.0411	73.68	2800	0.4673	0.2494
0.0242	78.94	3000	0.4801	0.2426
0.0833	84.21	3200	0.4641	0.2630
0.034	89.47	3400	0.4607	0.2449
0.02	94.73	3600	0.4825	0.2449
0.0211	99.99	3800	0.4822	0.2449

Framework versions

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