sammy786/wav2vec2-xlsr-dhivehi

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

• Loss: 14.86
• Wer: 29.32

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	4.883800	3.190218	1.000000
400	1.600100	0.497887	0.726159
600	0.928500	0.358781	0.603892
800	0.867900	0.309132	0.570786
1000	0.743100	0.309116	0.552954
1200	0.725100	0.266839	0.538378
1400	0.786200	0.259797	0.535897
1600	0.655700	0.245691	0.517290
1800	0.650500	0.246957	0.516204
2000	0.685500	0.234808	0.516204
2200	0.487100	0.228409	0.507753
2400	0.401300	0.221087	0.495968
2600	0.359300	0.212476	0.489301
2800	0.347300	0.204848	0.487750
3000	0.327000	0.203163	0.478756
3200	0.337100	0.210235	0.487595
3400	0.308900	0.201471	0.491316
3600	0.292600	0.192437	0.476120
3800	0.289600	0.198398	0.468445
4000	0.290200	0.193484	0.467204
4200	0.272600	0.193999	0.470150
4400	0.266700	0.187384	0.460769
4600	0.253800	0.187279	0.476663
4800	0.266400	0.197395	0.466817
5000	0.258000	0.188920	0.456660
5200	0.237200	0.180770	0.457358
5400	0.237900	0.178149	0.448287
5600	0.232600	0.179827	0.461002
5800	0.228500	0.182142	0.445185
6000	0.221000	0.173619	0.440688
6200	0.219500	0.172291	0.442859
6400	0.219400	0.173339	0.430609
6600	0.201900	0.177552	0.426423
6800	0.199000	0.173157	0.429834
7000	0.200000	0.166503	0.423709
7200	0.194600	0.171812	0.429834
7400	0.192100	0.164989	0.420530
7600	0.185000	0.168355	0.418825
7800	0.175100	0.168128	0.419290
8000	0.173500	0.167959	0.424950
8200	0.172200	0.173643	0.414793
8400	0.164200	0.167020	0.406342
8600	0.170800	0.168050	0.405334
8800	0.157900	0.164290	0.396573
9000	0.159900	0.163188	0.397426
9200	0.151700	0.164370	0.390991
9400	0.146600	0.165053	0.392852
9600	0.142200	0.164939	0.391844
9800	0.148300	0.164422	0.385719
10000	0.136200	0.166569	0.385951
10200	0.140700	0.161377	0.379594
10400	0.133300	0.165194	0.378276
10600	0.131300	0.164328	0.369205
10800	0.135500	0.160254	0.373236
11000	0.121100	0.163522	0.372693

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-dhivehi --dataset mozilla-foundation/common_voice_8_0 --config dv --split test