sammy786/wav2vec2-xlsr-chuvash

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

• Loss: 18.02
• Wer: 29.22

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	6.559100	2.274687	1.000000
400	1.346100	0.508268	0.681995
600	0.797500	0.391174	0.572876
800	0.556300	0.308620	0.489283
1000	0.435800	0.273956	0.454014
1200	0.388700	0.311027	0.499415
1400	0.338300	0.243977	0.413874
1600	0.294000	0.214134	0.385230
1800	0.276000	0.245991	0.397311
2000	0.253900	0.208324	0.363016
2200	0.233600	0.222156	0.370811
2400	0.219700	0.202602	0.364186
2600	0.205000	0.241339	0.384451
2800	0.176000	0.263558	0.384061
3000	0.166700	0.211768	0.333398
3200	0.160600	0.198677	0.321512
3400	0.154600	0.208655	0.328722
3600	0.146800	0.188022	0.317810
3800	0.133200	0.181083	0.313133
4000	0.134200	0.190084	0.316251
4200	0.114200	0.193034	0.312159
4400	0.117300	0.194122	0.312354
4600	0.112300	0.191111	0.305534
4800	0.107800	0.185930	0.302611
5000	0.100400	0.178625	0.299883
5200	0.099800	0.176442	0.294622
5400	0.100800	0.177935	0.294427
5600	0.096300	0.182903	0.293843
5800	0.094200	0.181041	0.293453
6000	0.097600	0.179865	0.290725
6200	0.091600	0.180327	0.292868
6400	0.093100	0.180275	0.292284

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-chuvash --dataset mozilla-foundation/common_voice_8_0 --config cv --split test