ayameRushia/wav2vec2-large-xls-r-300m-mn

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

Training Loss	Epoch	Step	Validation Loss	Wer
No log	6.35	400	0.9380	0.7902
3.2674	12.7	800	0.5794	0.5309
0.7531	19.05	1200	0.5749	0.4815
0.5382	25.4	1600	0.5530	0.4447
0.4293	31.75	2000	0.5709	0.4237
0.4293	38.1	2400	0.5476	0.4059

Framework versions

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