W2V2 BERT Withlm Malayalam

🚀 W2V2-BERT-withLM-Malayalam

This model is a fine - tuned version of facebook/w2v-bert-2.0 on multiple datasets, including IMASC, MSC, OpenSLR Malayalam Train split, Festvox Malayalam, and CV16. It is designed for automatic speech recognition, aiming to accurately transcribe Malayalam speech.

🚀 Quick Start

This model is ready to use for automatic speech recognition. You can load it using the Hugging Face Transformers library and start transcribing Malayalam audio.

✨ Features

• Fine - tuned Model: Based on facebook/w2v-bert-2.0, fine - tuned on multiple Malayalam datasets for better performance.
• Trigram Language Model: Trained using the KENLM Library on [kavyamanohar/ml - sentences](https://huggingface.co/datasets/kavyamanohar/ml - sentences) dataset to improve transcription accuracy.

📚 Documentation

Model Information

Model Performance

It achieves the following results on the validation set OpenSLR - Test:

• Loss: 0.1722
• Wer: 0.1299

Model Index

• Name: W2V2 - BERT - withLM - Malayalam by Bajiyo Baiju, Kavya Manohar
• Results:
  • Task: Automatic Speech Recognition
  • Datasets and Metrics:
    • OpenSLR Malayalam - Test: Wer = 18.23
    • Google Fleurs: Wer = 31.92
    • Mozilla Common Voice: Wer = 49.79

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e - 05
• 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: 10
• mixed_precision_training: Native AMP

Training Results

Framework Versions

• Transformers 4.39.3
• Pytorch 2.1.1+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1

📄 License

This model is released under the MIT license.