OpenTalking

What OpenTalking is

OpenTalking is a free, Apache-2.0 framework for real-time conversations with digital-human avatars. It connects frontend interaction, LLM responses, speech services, and avatar video playback. Its workflow includes session state, speech recognition and synthesis, voice selection, interruption control, subtitle events, and WebRTC audio and video. Pluggable model backends include Mock, QuickTalk, Wav2Lip, and remote inference options such as FlashTalk. OpenAI-compatible endpoints are supported, with Qwen, DeepSeek, and GPT documented as examples. Deployment can be local, private, Docker-based, or distributed, with API, worker, frontend, and model services separated. Mock mode checks API and LLM connections, streaming text-to-speech, subtitle events, WebRTC delivery, and browser playback without downloading video model weights. Linux with CUDA 12 is the main deployment target; macOS supports Mock mode only, and Windows uses Mock mode with WSL2 recommended. A WebRTC-capable client is required, and Safari needs extra CORS configuration. Example scenarios include commerce, tourism, medical guidance, broadcasting, companionship, and character content.