Voicebox Android

Voicebox Android

Offline Voice Cloning on Android

AIOfficial site

Description

Cloning a voice with GPT-SoVITS used to mean a desktop with a decent GPU and an evening lost to Python environments. Voicebox runs it on an Android phone instead: record about five seconds of reference audio, type its transcript, and the app speaks any text in that voice, fully offline once the models are downloaded, with no cloud and no graphics card.

It is quick, too. On a Xiaomi 14 with a Snapdragon 8 Gen 3, the V2 Pro Plus model produces a four-second sentence in about 2.6 seconds and runs roughly twice real-time on long passages. Free and open source, it needs an ARM64 phone on Android 11 or later and speaks Chinese, Japanese and English.

Features



Cross-lingual cloning: The reference clip and output language are independent, so a voice cloned from Japanese audio can speak Chinese or English.

Bring your own weights: Import GPT-SoVITS .ckpt and .pth checkpoints and convert them on the phone, with support for V2 Pro, V2 Pro Plus and V5 Turbo.

GPU and NPU acceleration: Inference runs on WebGPU, and compatible Snapdragon devices can add the NPU so even V5 Turbo beats real-time on long text.

QQ voice messages: Send results as native tap-to-play QQ voice notes rather than audio files, up to 60 seconds each.

Local HTTP API: A built-in server lets Tasker, audiobook readers and other apps request speech, and a PC can reach it over ADB port forwarding.