Real-Time AI Voice Conversion App

A cross-platform app that converts speech into selected voices in real time while integrating with existing calling and streaming applications.

Challenge

Natural voice conversion for gaming, streaming, and online meetings required both high audio quality and latency low enough not to interrupt conversation. The app also needed to support different iOS and Android audio pipelines and switch voice presets without interrupting a call.

Solution

Compared multiple voice conversion models and selected RVC for its balance of real-time performance and quality. Applied INT8 quantization, inference optimization, and persistent encoder and vocoder instances, then connected React Native with native audio modules.

Result

Delivered perceived latency below 200ms and released a shared iOS and Android app that works with major calling and streaming applications.

Team

2 members, 5 months

Frontend, audio pipeline, and AI model evaluation

Role

Handled cross-platform frontend development and audio pipeline optimization.

Implemented the React Native UI, evaluated AI models, optimized inference, and integrated native audio processing on iOS and Android.

Tech Stack

FrontendReact Native / TypeScript
AI ModelRVC / So-VITS-SVC / DDSP-SVC
RuntimeONNX Runtime / TorchScript
AudioAVAudioEngine / AAudio / Oboe
BackendPython / FastAPI

Key Features

01

Real-time AI voice conversion with perceived latency below 200ms

02

Voice preset switching without reloading the full model

03

Shared React Native codebase for iOS and Android

04

Integration with major calling and streaming applications

Technical Highlights

Model Selection for Real-Time Use

Compared RVC, So-VITS-SVC, and DDSP-SVC, then selected RVC based on practical latency and audio quality.

Mobile Inference Optimization

Combined INT8 quantization, warm caches, and persistent model components to keep inference stable on mobile devices.