AI product Open source
Sophon-Realtime is an open-source family of full-duplex voice models from Adventists AI. It continuously listens while a person speaks, uses a speech model to decide whether the speaker has finished, is back-channeling, or is interrupting, and drafts a reply from pauses before releasing audio. Its reply can include expressive sound events such as laughter, sighs, breaths, coughs, and emphasis; barge-in handling discards an interrupted draft and continues with the new question.
The family includes Tag, Flash, and E2E builds. Tag and Flash generate sound-tagged replies that an expressive speech model performs, while E2E generates continuous voice vectors directly with a compact talker and decoder. The models support speech-to-speech interaction, tool calling, OpenAI Realtime-compatible interfaces, zero-shot voice cloning from a short clip, and a Live mode that delegates lookups and actions to a user's backend. They can run locally with vLLM-Omni on CUDA or with MLX on Apple silicon, or be accessed through Adventists AI's hosted API. Code and weights are distributed under Apache-2.0.
1 use taken from transcripts — each links to the moment in the video.
An open-source family of full-duplex voice models from Adventist AI that hear your tone and reply with expressive speech (laughs, sighs, breaths), reaching first audio in ~0.3 seconds with tool calling, built in Python on CUDA and served with vLLM under Apache 2.0.
1 in the library.