wav2vec2-xls-r-300m-ca

wav2vec2-xls-r-300m-ca

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - CA, the tv3_parla and parlament_parla datasets. It achieves the following results on the evaluation set (for the three datasets):

• Loss: 0.2472
• Wer: 0.1499

Model description

Please check the original facebook/wav2vec2-xls-r-1b Model card. This is just a finetuned version of that model.

Intended uses & limitations

As any model trained on crowdsourced data, this model can show the biases and particularities of the data and model used to train this model. Moreover, since this is a speech recognition model, it may underperform for some lower-resourced dialects for the catalan language.

Training and evaluation data

More information needed

Training procedure

The data is preprocessed to remove characters not on the catalan alphabet. Moreover, numbers are verbalized using code provided by @ccoreilly, which can be found on the text/ folder or here.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 18.0
• mixed_precision_training: Native AMP

Training results

Check the Tensorboard tab to check the training profile and evaluation results along training. The model was evaluated on the test splits for each of the datasets used during training.

Framework versions

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

Thanks

Want to thank both @ccoreilly and @gullabi who have contributed with their own resources and knowledge into making this model possible.