sammy786/wav2vec2-xlsr-kyrgyz

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

• Loss: 43.06
• Wer: 39.19

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: 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	5.357800	2.700367	1.000000
400	1.513600	0.642542	0.598820
600	0.961900	0.530665	0.502739
800	0.776000	0.507709	0.462705
1000	0.646100	0.453115	0.444164
1200	0.581200	0.454797	0.438264
1400	0.437900	0.459389	0.426464
1600	0.348600	0.401247	0.416351
1800	0.312800	0.436135	0.409608
2000	0.294100	0.440911	0.398651
2200	0.281400	0.432729	0.394016
2400	0.258400	0.429860	0.393595
2600	0.263700	0.432689	0.395280
2800	0.256900	0.430672	0.391909

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-kyrgyz --dataset mozilla-foundation/common_voice_8_0 --config ky --split test