An SDK for real-time avatars that never phones home

What it does

Duix Mobile is a C++ SDK that renders real-time, talking AI avatars on phones, tablets, automotive screens, and embedded displays. You supply your own LLM, speech recognition, and text-to-speech services; the SDK handles the lip-synced facial animation, expression changes, and streaming audio entirely on-device. The README pitches it for customer-service bots, virtual tutors, and—yes—virtual companions.

The interesting bit

The whole pipeline—animation, audio synthesis, and interruption handling—runs locally, which is why the project claims sub-120 ms response latency on a Snapdragon 8 Gen 2. That local-first design is pitched specifically at finance, government, and legal scenarios where keeping data off the network is a requirement, not a preference.

Key highlights

• Fully offline rendering: core avatar animation and audio run on-device with no cloud dependency.
• BYO backend: built to integrate custom or third-party LLM, ASR, and TTS services.
• Streaming audio with barge-in: avatars can be interrupted mid-sentence, and voice start/end callbacks are documented.
• Cross-platform C++ core targeting Android, iOS, tablets, automotive, and VR/IoT displays.
• Four pre-built avatar models (Leo, Oliver, Sofia, Lily) ship as downloadable assets.

Caveats

• Custom avatars require emailing the vendor; the README says a 15-second to 2-minute video is “typically sufficient,” but pricing and turnaround are unspecified.
• The repo description claims “<1.5 s latency” while the README claims “under 120ms” on a specific SoC; the discrepancy is unexplained.

Verdict

Worth evaluating if you need a local-only avatar renderer for kiosks, tablets, or mobile apps in regulated industries. Skip it if you were hoping for a fully open-source pipeline that bundles its own LLM, ASR, or TTS models out of the box.