wav2vec2-large-xls-r-300m-assamese

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_7_0 dataset. It achieves the following results on the evaluation set:

• WER: 0.7954545454545454
• CER: 0.32341269841269843

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

To compute the evaluation parameters

cd wav2vec2-large-xls-r-300m-assamese; python eval.py --model_id ./ --dataset mozilla-foundation/common_voice_7_0 --config as --split test --log_outputs

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-4
• train_batch_size: 16
• eval_batch_size: 8
• seed: not given
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 400
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.584065	NA	400	1.584065	0.915512
1.658865	Na	800	1.658865	0.805096
1.882352	NA	1200	1.882352	0.820742
1.881240	NA	1600	1.881240	0.810907
2.159748	NA	2000	2.159748	0.804202
1.992871	NA	2400	1.992871	0.803308
2.201436	NA	2800	2.201436	0.802861
2.165218	NA	3200	2.165218	0.793920
2.253643	NA	3600	2.253643	0.796603
2.265880	NA	4000	2.265880	0.790344
2.293935	NA	4400	2.293935	0.797050
2.288851	NA	4800	2.288851	0.784086

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu113
• Datasets 1.13.3
• Tokenizers 0.10.3