This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the Common Voice 8 dataset of five similar languages with similar scripts: Czech, Slovak, Polish, Slovenian and Upper Sorbian. Training and validation sets were concatenated and shuffled.

Evaluation set used for training was concatenated from the respective test sets and shuffled while limiting each language to at most 2000 samples. During training, cca WER 70 was achieved on this set.

Evaluation script

python eval.py --model_id comodoro/wav2vec2-xls-r-300m-west-slavic-cv8 --dataset mozilla-foundation/common_voice_8_0 --split test --config {lang}

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 32
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 50
mixed_precision_training: Native AMP

Framework versions

Transformers 4.16.0.dev0
Pytorch 1.10.1+cu102
Datasets 1.18.3
Tokenizers 0.11.0