AigizK/wav2vec2-large-xls-r-300m-bashkir-cv7_opt

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

wav2vec2-large-xls-r-300m-bashkir-cv7_opt

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMONVOICE7_0 - BA dataset. It achieves the following results on the evaluation set:

• —Training Loss: 0.268400
• —Validation Loss: 0.088252
• —WER without LM: 0.085588
• —WER with LM: 0.04440795062008041
• —CER with LM: 0.010491234992390509

Model description

Trained with this jupiter notebook

Intended uses & limitations

In order to reduce the number of characters, the following letters have been replaced or removed:

• —'я' -> 'йа'
• —'ю' -> 'йу'
• —'ё' -> 'йо'
• —'е' -> 'йэ' for first letter
• —'е' -> 'э' for other cases
• —'ъ' -> deleted
• —'ь' -> deleted

Therefore, in order to get the correct text, you need to do the reverse transformation and use the language model.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0001
• —trainbatchsize: 32
• —evalbatchsize: 32
• —seed: 42
• —gradientaccumulationsteps: 2
• —totaltrainbatch_size: 64
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: linear
• —lrschedulerwarmup_steps: 300
• —num_epochs: 50
• —mixedprecisiontraining: Native AMP

Framework versions

• —Transformers 4.16.1
• —Pytorch 1.10.0+cu113
• —Datasets 1.18.2
• —Tokenizers 0.10.3