~/AI/adventists-ai-releases-sophon-realtime-v1-open-source-voice-models

Adventists AI Releases Sophon-Realtime-v1 Open-Source Voice Models

Adventists AI has released Sophon-Realtime-v1, a new open-source family of full-duplex voice models that natively handle tone, interruptions, and expressive non-verbal cues. The family includes three builds—Tag, Flash, and E2E—all powered by a shared 32B model that listens, decides turn-taking, and generates responses. This release represents a major milestone for open-source conversational AI, offering performance comparable to top proprietary models like GPT-4o-mini-Audio on benchmarks while running locally. It enables developers to build low-latency, emotionally expressive voice applications with commercial-friendly Apache-2.0 licensing. The Tag build achieves top scores on VoiceBench (82.8%) and paralinguistic understanding, delivering first-audio latency of about 0.6 seconds and supporting up to 16 concurrent users on an 80 GB GPU via vLLM-Omni. Meanwhile, the Flash build prioritizes speed, dropping first-audio latency to 0.31 seconds and supporting up to 40 users per GPU.

## BACKGROUND

Full-duplex voice models utilize continuous audio streaming architectures that allow a system to listen and speak simultaneously without waiting for the user to finish talking completely. This enables natural conversational dynamics, including immediate interruptions and real-time adjustments based on vocal tones and non-verbal sounds.

## KEYWORDS

#AI#Voice Models#Open Source#LocalLLaMA#Speech Synthesis

$ subscribe --daily