wav2vec2-large-xls-r-300m-dansk-CV-80

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m for Danish, using the mozilla-foundation/common_voice_8_0 dataset.

It achieves the following results on the evaluation set:

• eval_loss: 0.6394
• eval_wer: 0.3682
• eval_runtime: 104.0466
• eval_samples_per_second: 13.359
• eval_steps_per_second: 1.672
• epoch: 21.28
• step: 2000

Model description

ASR Danish model

Intended uses & limitations

More information needed

Training and evaluation data

Danish subset of mozilla-foundation/common_voice_8_0

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• 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: 30
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.16.1
• Pytorch 1.10.0+cu111
• Datasets 1.18.2
• Tokenizers 0.11.0