Oriel — AI & technology worth noticing, every day.

adventists-ai/Sophon-Realtime-v1

AI model · https://huggingface.co/adventists-ai/Sophon-Realtime-v1

First seen by Oriel on .

AI summary

Sophon-Realtime v1 is an open Apache-2.0 full-duplex voice model family that handles interruptions and turn-taking while interpreting vocal tone and producing expressive speech.

  • Its three builds share a 32B model: Tag prioritizes expressiveness, Flash supports up to 40 concurrent users per 80 GB GPU, and E2E generates speech with a compact 350M-parameter talker.
  • The model listens continuously, drafts replies during pauses, drops drafts when the user continues, and retains only audio the user actually heard after an interruption.
  • Reported scores include VoiceBench 82.8, Big Bench Audio 81.8%, paralinguistic understanding 48.6, and tool-calling Pass@1 of 48% with a 100% response rate.
  • Tag and Flash use vLLM-Omni and share APIs; E2E uses its own engine. First audio is reported at about 0.6 seconds for Tag, 0.31 seconds for Flash, and 0.25 seconds for E2E.
  • Developers can serve the models through realtime and chat-completions APIs, use a GPT-Live-compatible gateway, or fine-tune the LLM backbone with LoRA using ms-swift.

#speech models #full-duplex audio #voice synthesis #model fine-tuning

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 344, latest observation ; +255 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 0, latest observation ; 0 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsadventists-ai/Sophon-Realtime-v1https://huggingface.co/adventists-ai/Sophon-Realtime-v1 Read discussion