Kabyle Xlsr

🚀 Akashpb13/Kabyle_xlsr

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA - FOUNDATION/COMMON_VOICE_7_0 - hu dataset. It offers a practical solution for automatic speech recognition tasks, leveraging the pre - trained capabilities of the base model and adapting them to specific language data.

📚 Documentation

Model description

"facebook/wav2vec2-xls-r-300m" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Kabyle train.tsv. Only 50,000 records were sampled randomly and trained due to the huge size of the dataset. Only those points were considered where upvotes were greater than downvotes and duplicates were removed after concatenation of all the datasets given in common voice 7.0

Training procedure

For creating the training dataset, all possible datasets were appended and a 90 - 10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.000096
• train_batch_size: 8
• seed: 13
• gradient_accumulation_steps: 4
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.0+cu102
• Datasets 1.18.3
• Tokenizers 0.10.3

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id Akashpb13/Kabyle_xlsr --dataset mozilla-foundation/common_voice_8_0 --config kab --split test

📄 License

The model is released under the Apache - 2.0 license.

🔍 Model Index

This model achieves the following results on the evaluation set (which is 10 percent of the train dataset merged with dev datasets):

• Loss: 0.159032
• Wer: 0.187934