sammy786/wav2vec2-xlsr-estonian

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

• Loss: 17.94
• Wer: 30.38

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: 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: 30
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	3.729100	1.096018	0.959867
400	0.996900	0.310228	0.443600
600	0.762900	0.210873	0.346117
800	0.621400	0.200381	0.331513
1000	0.408000	0.196382	0.322014
1200	0.320200	0.176281	0.312515
1400	0.315300	0.179433	0.303847
1600	0.445800	0.420985	0.315839
1800	0.644600	0.433833	0.354904
2000	0.550900	0.327117	0.336500
2200	0.498600	0.289830	0.325457
2400	0.488300	0.294309	0.314177
2600	0.491700	0.311175	0.318689
2800	0.508500	0.314744	0.320470
3000	0.499900	0.314834	0.320589

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-estonian --dataset mozilla-foundation/common_voice_8_0 --config et --split test