sammy786/wav2vec2-xlsr-romansh_sursilvan

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

Loss: 16.38
Wer: 21.25

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	4.825500	2.932350	1.000000
400	1.325600	0.292645	0.415436
600	0.709800	0.219167	0.324451
800	0.576800	0.174390	0.275477
1000	0.538100	0.183737	0.272116
1200	0.475200	0.159078	0.253871
1400	0.420400	0.167277	0.240907
1600	0.393500	0.167216	0.247269
1800	0.407500	0.178282	0.239827
2000	0.374400	0.184590	0.239467
2200	0.382600	0.164106	0.227824
2400	0.363100	0.162543	0.228544
2600	0.199000	0.172903	0.231665
2800	0.150800	0.160117	0.222662
3000	0.101100	0.169553	0.222662
3200	0.104200	0.161056	0.220622
3400	0.096900	0.161562	0.216781
3600	0.092200	0.163880	0.212580
3800	0.089200	0.162288	0.214140
4000	0.076200	0.160470	0.213540
4200	0.087900	0.162827	0.213060
4400	0.066200	0.161096	0.213300
4600	0.076000	0.162060	0.213660
4800	0.071400	0.162045	0.213300

Framework versions

Transformers 4.16.0.dev0
Pytorch 1.10.0+cu102
Datasets 1.17.1.dev0
Tokenizers 0.10.3

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id sammy786/wav2vec2-xlsr-romansh_sursilvan --dataset mozilla-foundation/common_voice_8_0 --config rm-sursilv --split test