sammy786/wav2vec2-xlsr-basaa

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

• Loss: 21.39
• Wer: 30.99

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: 16
• eval_batch_size: 16
• seed: 13
• gradient_accumulation_steps: 2
• 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: 70
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	6.734100	1.605006	0.980456
400	1.011200	0.364686	0.442997
600	0.709300	0.300204	0.377850
800	0.469800	0.315612	0.405537
1000	0.464700	0.352494	0.372964
1200	0.421900	0.342533	0.368078
1400	0.401900	0.351398	0.343648
1600	0.429800	0.350570	0.348534
1800	0.352600	0.356601	0.358306
2000	0.387200	0.355814	0.356678
2200	0.362400	0.345573	0.355049

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-basaa --dataset mozilla-foundation/common_voice_8_0 --config bas --split test