Show HN: SPEV: Open Weight, Omnimodal Decision Model Trained for Turn Detection

Category: ai-ml

Tags: omnimodal, turn-detection, voice-ai, decision-model, conversational-ai

Score: 7.0/10 (Innovation: 7, Technical: 8, Documentation: 5, Utility: 6)

SPEV is a 30B parameter omnimodal decision model built on Qwen3-Omni, fine-tuned for turn detection in conversational audio. It evaluates multiple-choice questions over text, audio, images, and video, returning per-choice probabilities via a custom evidence-attention head. The model addresses a specific gap in voice AI by providing a reusable component for deciding when a speaker has yielded the floor.

Target audience: AI/ML engineers building voice assistants and conversational agents

Repository: https://huggingface.co/SpragAI/SPEV

View on Hacker News