sammy786/wav2vec2-xlsr-mongolian

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

• Loss: 31.52
• Wer: 34.1522

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: 30
• mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
200	4.906200	3.012986	1.000000
400	1.734600	0.704821	0.750497
600	1.132100	0.496223	0.531241
800	0.929300	0.468937	0.469043
1000	0.772300	0.425313	0.448168
1200	0.623900	0.394633	0.414229
1400	0.512400	0.369225	0.397614
1600	0.439900	0.346033	0.391650
1800	0.391300	0.358454	0.379296
2000	0.377000	0.346822	0.359415
2200	0.347500	0.325205	0.348481
2400	0.343600	0.315233	0.344078
2600	0.328000	0.308826	0.341522
2800	0.358200	0.331786	0.343084
3000	0.417200	0.370051	0.356433
3200	0.685300	0.595438	0.407413
3400	0.764100	0.643449	0.359983
3600	0.717100	0.505033	0.371911
3800	0.620900	0.464138	0.369071
4000	0.590700	0.445417	0.363249
4200	0.561000	0.440727	0.360267
4400	0.550600	0.447122	0.360267
4600	0.562100	0.457020	0.359841
4800	0.578800	0.470477	0.360551
5000	0.580400	0.481413	0.362539
5200	0.605500	0.485240	0.362823
5400	0.582900	0.486654	0.362965
5600	0.593900	0.486715	0.363107
5800	0.590900	0.486716	0.363107
6000	0.587200	0.486716	0.363107

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-mongolian --dataset mozilla-foundation/common_voice_8_0 --config mn --split test