IndicVoiceChanger
IndicVoiceChanger is a finetuned version of the Seed Voice Conversion model, adapted for Indian languages.
It enables high-quality voice conversion across multiple Indian languages, preserving speaker identity while changing the voice characteristics.
Overview
This model is built upon the Seed Voice Conversion checkpoints and finetuned with a mix of publicly available open-source datasets and our own proprietary dataset.
It is designed to work well on speech data from diverse Indian languages, accents, and speaking styles.
Try It Out
Experience the model firsthand at: Hugging Face Spaces Demo
Fine-tuning on Custom Data
While the zero shot performance of this model is usually good, it can be further improved with fine-tuning. This model supports efficient fine-tuning on your custom speakers with remarkable data efficiency and speed:
- Minimal Data Requirements: Train on new speakers with as little as 1 utterance per speaker
- Ultra-Fast Training: Achieve good results in just 100 training steps (approximately 2 minutes on T4 GPU)
- Speaker Adaptation: Significantly improve performance on specific target speakers through personalized fine-tuning
Getting Started with Fine-tuning
For detailed instructions on installation, usage, and fine-tuning, please refer to the comprehensive guide at: https://github.com/Plachtaa/seed-vc
Note: We will be updating this repository with example notebooks demonstrating the fine-tuning process soon.
Important: When fine-tuning for Indian languages, make sure to use the checkpoints provided in this repository as your starting point for optimal performance.
Acknowledgments
Special thanks to the SeedVC project for providing the foundational architecture and training framework that made this Indian language adaptation possible.