This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - MN dataset. It achieves the following results on the evaluation set:

• Loss: 0.5502
• Wer: 0.4042

Training and evaluation data

Evaluation is conducted in Notebook, you can see within the repo "notebook_evaluation_wav2vec2_mn.ipynb"

Test WER without LM wer = 58.2171 % cer = 16.0670 %

Test WER using wer = 31.3919 % cer = 10.2565 %

How to use eval.py

huggingface-cli login #login to huggingface for getting auth token to access the common voice v8
#running with LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-mn --dataset mozilla-foundation/common_voice_8_0 --config mn --split test

# running without LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-mn --dataset mozilla-foundation/common_voice_8_0 --config mn --split test --greedy

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 40.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0