Lemswasabi/wav2vec2-large-xlsr-53-842h-luxembourgish-11h-with-lm

Model description

We fine-tuned a wav2vec 2.0 large XLSR-53 checkpoint with 842h of unlabelled Luxembourgish speech collected from RTL.lu. Then the model was fine-tuned on 11h of labelled Luxembourgish speech from the same domain. Additionally, we rescore the output transcription with a 5-gram language model trained on text corpora from RTL.lu and the Luxembourgish parliament.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.5e-05
• train_batch_size: 3
• eval_batch_size: 3
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 12
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.20.0.dev0
• Pytorch 1.11.0+cu113
• Datasets 2.2.1
• Tokenizers 0.12.1

Citation

This model is a result of our paper IMPROVING LUXEMBOURGISH SPEECH RECOGNITION WITH CROSS-LINGUAL SPEECH REPRESENTATIONS submitted to the IEEE SLT 2022 workshop

@misc{lb-wav2vec2,
  author = {Nguyen, Le Minh and Nayak, Shekhar and Coler, Matt.},
  keywords = {Luxembourgish, multilingual speech recognition, language modelling, wav2vec 2.0 XLSR-53, under-resourced language},
  title = {IMPROVING LUXEMBOURGISH SPEECH RECOGNITION WITH CROSS-LINGUAL SPEECH REPRESENTATIONS},
  year = {2022},
  copyright = {2023 IEEE}
}