Wav2vec2 Large Xlsr 53 Kalmyk

🚀 Wav2Vec2 Model for Kalmyk Speech Recognition

This Wav2Vec2 model is designed for Kalmyk speech recognition, leveraging pre - training on extensive datasets and fine - tuning for better performance.

🚀 Quick Start

This model is ready for use in Kalmyk speech recognition tasks. You can integrate it into your projects following the general steps of using Wav2Vec2 models.

✨ Features

• Rich Pretraining: The model was initially pretrained on 500 hours of Kalmyk TV recordings and 1000 hours of Mongolian speech recognition dataset, providing a solid foundation for speech understanding.
• Fine - tuning on Synthetic Data: It was then finetuned on a 300 - hour Kalmyk synthetic STT dataset created by a voice conversion model, which helps adapt to specific Kalmyk speech characteristics.
• Performance Metrics:
  • Achieved a 50% Word Error Rate (WER) on a private test set created from Kalmyk TV recordings.
  • Expected to have a much lower WER on clean voice recordings.

🔧 Technical Details

Voice Conversion Info

• Synthetic Dataset: A 300 - hour Kalmyk synthetic STT dataset was used for fine - tuning.
• Source Voice: The source voice is a Kalmyk female voice TTS.
• Target Voices: Target voices are from the VCTK dataset.
• Example Data: You can find example data at https://twitter.com/tugstugi/status/1409111296897912835.
• Data Content: Each WAV file in the dataset has a different text created from Kalmyk books.

📄 License

This project is licensed under the Apache - 2.0 license.