sammy786/wav2vec2-xlsr-georgian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - ka dataset. It achieves the following results on evaluation set (which is 10 percent of train data set merged with other and dev datasets):

• Loss: 10.54
• Wer: 27.53

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
• train_batch_size: 8
• eval_batch_size: 16
• seed: 13
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	4.152100	0.823672	0.967814
400	0.889500	0.196740	0.444792
600	0.493700	0.155659	0.366115
800	0.328000	0.138066	0.358069
1000	0.260600	0.119236	0.324989
1200	0.217200	0.114050	0.313366
1400	0.188800	0.112600	0.302190
1600	0.166900	0.111154	0.295485
1800	0.155500	0.109963	0.286544
2000	0.140400	0.107587	0.277604
2200	0.142600	0.105662	0.277157
2400	0.135400	0.105414	0.275369

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

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