Wav2vec2 Large Xls R 300m Ia

🚀 wav2vec2-large-xls-r-300m-ia

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It is designed for automatic speech recognition tasks and has achieved remarkable results on the evaluation set.

🚀 Quick Start

Evaluation

You can evaluate the model using the following commands:

huggingface-cli login #login to huggingface for getting auth token to access the common voice v8
#running with LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-ia --dataset mozilla-foundation/common_voice_8_0 --config ia --split test

# running without LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-ia --dataset mozilla-foundation/common_voice_8_0 --config ia --split test --greedy

✨ Features

• Fine - tuned Model: Based on facebook/wav2vec2-xls-r-300m, fine - tuned on the common_voice dataset.
• Good Performance: Achieves low WER (Word Error Rate) and CER (Character Error Rate) on the evaluation set.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

📚 Documentation

Model Information

Evaluation Results

This model achieves the following results on the evaluation set:

• Loss: 0.1452
• Wer: 0.1253

Training Procedure

Training is conducted in Google Colab, and the training notebook is provided in the repo.

Training and Evaluation Data

• Language Model: Created from texts from processed sentences in the train + validation split of the dataset (common voice 8.0 for Interlingua).
• Evaluation: Conducted in the notebook "notebook_evaluation_wav2vec2_ia.ipynb" within the repo.

Evaluation Metrics

• Test WER without LM:
  • wer = 20.1776 %
  • cer = 4.7205 %
• Test WER using LM:
  • wer = 8.6074 %
  • cer = 2.4147 %

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e - 05
• train_batch_size: 16
• eval_batch_size: 4
• 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: 400
• num_epochs: 30
• mixed_precision_training: Native AMP

Training Results

Framework Versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.11.0

📄 License

This project is licensed under the Apache - 2.0 license.