Xls Asr Vi 40h 1B

🚀 xls-asr-vi-40h-1B

This model is a fine - tuned version of facebook/wav2vec2-xls-r-1b on 40 hours of FPT Open Speech Dataset (FOSD) and Common Voice 7.0. It is designed for automatic speech recognition tasks, leveraging data from Common Voice and other relevant datasets to enhance performance.

✨ Features

• Fine - Tuned Model: Based on facebook/wav2vec2-xls-r-1b, fine - tuned on specific datasets for better performance in Vietnamese speech recognition.
• Benchmark Results: Provides clear WER and CER results on multiple datasets, including VIVOS, Common Voice 7.0, and Common Voice 8.0, with and without Language Models.

📚 Documentation

Model Information

Benchmark Results

Benchmark WER result

Benchmark CER result

Model Index

• Name: xls - asr - vi - 40h - 1B
• Results:
  • Task: Speech Recognition (automatic - speech - recognition)
    • Dataset: Common Voice 7.0 (mozilla - foundation/common_voice_7_0, args: vi)
      • Metrics:
        • Test WER (with LM): 25.846
        • Test CER (with LM): 12.961
    • Dataset: Common Voice 8.0 (mozilla - foundation/common_voice_8_0, args: vi)
      • Metrics:
        • Test WER (with LM): 31.158
        • Test CER (with LM): 16.179

💻 Usage Examples

Evaluation

Please use the eval.py file to run the evaluation.

python eval.py --model_id geninhu/xls-asr-vi-40h-1B --dataset mozilla-foundation/common_voice_7_0 --config vi --split test --log_outputs

🔧 Technical Details

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e - 05
• train_batch_size: 16
• eval_batch_size: 16
• 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: 1500
• num_epochs: 10.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0

📄 License

This project is licensed under the Apache - 2.0 license.