Wav2vec2 Large Xls R 300m Hsb V3

🚀 wav2vec2-large-xls-r-300m-hsb-v3

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - HSB dataset. It is designed for automatic speech recognition in the Upper Sorbian (hsb) language, and it has achieved certain results on the evaluation set, which makes it a valuable model in the field of speech recognition.

✨ Features

• Language Support: Specifically tailored for the Upper Sorbian (hsb) language, suitable for related speech recognition tasks.
• Fine - Tuned Model: Based on the pre - trained model facebook/wav2vec2-xls-r-300m, fine - tuned on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - HSB dataset to better adapt to the target language.
• Performance Metrics: Achieved specific results on the evaluation set, including Loss and Wer, which can be used as a reference for model performance.

📚 Documentation

Model Information

Model Results

The model has the following performance on different datasets:

• Common Voice 8:
  • Task: Automatic Speech Recognition
  • Metrics:
    • Test WER: 0.4763681592039801
    • Test CER: 0.11194945177476305
• Robust Speech Event - Dev Data:
  • Task: Automatic Speech Recognition
  • Metrics:
    • Test WER: NA
    • Test CER: NA

Evaluation Commands

Evaluate on mozilla - foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-hsb-v3 --dataset mozilla-foundation/common_voice_8_0 --config hsb --split test --log_outputs

Evaluate on speech - recognition - community - v2/dev_data

⚠️ Important Note

Upper Sorbian (hsb) language not found in speech - recognition - community - v2/dev_data!

Training Hyperparameters

The following hyperparameters were used during the training process:

• learning_rate: 0.00045
• 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: 50
• mixed_precision_training: Native AMP

Training Results

Framework Versions

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