Modelpublic

sammy786/wav2vec2-xlsr-interlingua

sourceHugging Faceapache-2.0updated 5y agoView on Hugging Face

0likes37downloads

Model Card

sammy786/wav2vec2-xlsr-interlingua

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMONVOICE8_0 - ia dataset. It achieves the following results on evaluation set (which is 10 percent of train data set merged with other and dev datasets):

—Loss: 5.44
—Wer: 19.78

Model description

"facebook/wav2vec2-xls-r-1b" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Finnish train.tsv, dev.tsv and other.tsv

Training procedure

For creating the train dataset, all possible datasets were appended and 90-10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

—learning_rate: 0.000045637994662983496
—trainbatchsize: 16
—evalbatchsize: 16
—seed: 13
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 32
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: cosinewithrestarts
—lrschedulerwarmup_steps: 500
—num_epochs: 30
—mixedprecisiontraining: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	4.649200	0.483339	0.511322
400	0.764700	0.133428	0.251288
600	0.563700	0.099292	0.227745
800	0.438800	0.087545	0.217445
1000	0.406800	0.072313	0.213848
1200	0.237500	0.066965	0.213766
1400	0.177800	0.064419	0.208126
1600	0.157100	0.065962	0.214011
1800	0.146600	0.059477	0.202076
2000	0.132800	0.055015	0.201831
2200	0.122000	0.055421	0.201749
2400	0.115700	0.054462	0.197826

Framework versions

—Transformers 4.16.0.dev0
—Pytorch 1.10.0+cu102
—Datasets 1.17.1.dev0
—Tokenizers 0.10.3

Evaluation Commands

1.To evaluate on mozilla-foundation/common_voice_8_0 with split test

bash

python eval.py --model_id sammy786/wav2vec2-xlsr-interlingua --dataset mozilla-foundation/common_voice_8_0 --config ia --split test