Speech-to-Speech Model: so-vits-svc
Overview
This repository contains a speech-to-speech model, specifically the so-vits-svc, trained to mimic the voice of Chamber, a character from the game Valorant. The model is designed for speech spoofing and voice conversion applications, offering a high level of accuracy compared to other models like the RVC model, which is faster but less precise.
Model Choice
The so-vits-svc model was chosen over the RVC model due to its superior accuracy, despite the latter's speed advantage. Future plans include training the model to detect and convert a variety of innovative voices, such as transforming songs into my own voice.
Model Details: so-vits-svc Model
The so-vits-svc model is a singing voice changer that uses ViTS (Variational Inference for Text-to-Speech). This model is particularly suited for high-quality voice conversion tasks, making it ideal for applications where accuracy is crucial. Here’s a brief overview of its components and functionality:
- ViTS (Variational Inference for Text-to-Speech): An advanced technique that combines variational inference methods with text-to-speech models, allowing for more nuanced and accurate voice conversions.
- Singing Voice Changer: Originally designed for changing singing voices, the model can adapt to various speech patterns, making it versatile for different applications, including character voice mimicking and real-time voice conversion.
Training Details
Dataset
- Character: Chamber from Valorant
- Voice Lines: Approximately 500 voice lines
- Source: Downloaded from a website providing mp3 files of Valorant character voices
Training Process
- Epochs: 2000
- Duration: Approximately 24 hours (including a 2-hour break)
- Hardware: RTX 3070 GPU with 8GB VRAM
Future Work
- The model will be further trained and experimented with to include more diverse and innovative voices. Possible applications include converting songs to a specified voice.
- Web Application
- Plans are underway to develop a web application that allows users to convert their voice to different characters in real-time. This application aims to provide a fun way to play games and prank friends by transforming voices into various character voices seamlessly.
Acknowledgments
- Valorant for providing the character and voice lines.
- Online resources for the mp3 files.
Contact
- Made with passion by Anirudh Sai Lanka.
- For any queries or contributions, please contact me at anirudh2002sai1234@gmail.com.
license: mit
- Downloads last month
- -