sammy786/wav2vec2-xlsr-tatar

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

• Loss: 7.66
• Wer: 7.08

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: 40
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	4.849400	1.874908	0.995232
400	1.105700	0.257292	0.367658
600	0.723000	0.181150	0.250513
800	0.660600	0.167009	0.226078
1000	0.568000	0.135090	0.177339
1200	0.721200	0.117469	0.166413
1400	0.416300	0.115142	0.153765
1600	0.346000	0.105782	0.153963
1800	0.279700	0.102452	0.146149
2000	0.273800	0.095818	0.128468
2200	0.252900	0.102302	0.133766
2400	0.255100	0.096592	0.121316
2600	0.229600	0.091263	0.124561
2800	0.213900	0.097748	0.125687
3000	0.210700	0.091244	0.125422
3200	0.202600	0.084076	0.106284
3400	0.200900	0.093809	0.113238
3600	0.192700	0.082918	0.108139
3800	0.182000	0.084487	0.103371
4000	0.167700	0.091847	0.104960
4200	0.183700	0.085223	0.103040
4400	0.174400	0.083862	0.100589
4600	0.163100	0.086493	0.099728
4800	0.162000	0.081734	0.097543
5000	0.153600	0.077223	0.092974
5200	0.153700	0.086217	0.090789
5400	0.140200	0.093256	0.100457
5600	0.142900	0.086903	0.097742
5800	0.131400	0.083068	0.095225
6000	0.126000	0.086642	0.091252
6200	0.135300	0.083387	0.091186
6400	0.126100	0.076479	0.086352
6600	0.127100	0.077868	0.086153
6800	0.118000	0.083878	0.087676
7000	0.117600	0.085779	0.091054
7200	0.113600	0.084197	0.084233
7400	0.112000	0.078688	0.081319
7600	0.110200	0.082534	0.086087
7800	0.106400	0.077245	0.080988
8000	0.102300	0.077497	0.079332
8200	0.109500	0.079083	0.088339
8400	0.095900	0.079721	0.077809
8600	0.094700	0.079078	0.079730
8800	0.097400	0.078785	0.079200
9000	0.093200	0.077445	0.077015
9200	0.088700	0.078207	0.076617
9400	0.087200	0.078982	0.076485
9600	0.089900	0.081209	0.076021
9800	0.081900	0.078158	0.075757
10000	0.080200	0.078074	0.074498
10200	0.085000	0.078830	0.073373
10400	0.080400	0.078144	0.073373
10600	0.078200	0.077163	0.073902
10800	0.080900	0.076394	0.072446
11000	0.080700	0.075955	0.071585
11200	0.076800	0.077031	0.072313
11400	0.076300	0.077401	0.072777
11600	0.076700	0.076613	0.071916
11800	0.076000	0.076672	0.071916
12000	0.077200	0.076490	0.070989
12200	0.076200	0.076688	0.070856
12400	0.074400	0.076780	0.071055
12600	0.076300	0.076768	0.071320
12800	0.077600	0.076727	0.071055
13000	0.077700	0.076714	0.071254

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-tatar --dataset mozilla-foundation/common_voice_8_0 --config tt --split test