live-talk

A static photo rarely feels like it is truly interacting with the listener — virtual avatars tend to be stiff, laggy, and emotionally flat. Live Talk turns any single photo into a virtual character that turns, smiles, and moves naturally in sync with the conversation, delivering low-latency, high-detail real-time responsiveness.

The core capability rests on a three-layer architecture: a dsh plugin shell handles publishing, installation, and UI; a dialogue character visualization core provides character manifests, a vendor-agnostic semantic pipeline (sentence splitting, emotion and action annotation, TTS cleaning), and a session event bridge; and a voice and video provider seam abstracts ASR, TTS, and avatar-media into swappable registries. Built-in providers include Doubao ASR and TTS, Jimeng, Volcano, and Vidu for real-time use, while the client side offers zero-key browser TTS and speech recognition so you can audition without any extra credentials.

It is well suited for developers building virtual assistants, stream companions, or playful interactive experiences on DeepSeek Harness, and for product teams who want a static image to speak, react, and sync emotions in real time.