ConvFill: a 0.6B talker reads a large reasoner's answers for fast, smart voice agents
futterneid · reddit · 2026-10-26
ConvFill, from UW UbiComp Lab students (presented at EMNLP), bridges the fast-vs-smart tradeoff in voice agents: a 0.6B talker model replies immediately to buy time while a large reasoner thinks, then reads out the reasoner's answer — like having an expert in your earbud at a panel. The author finds the core idea promising though the talker needs to be smarter. A demo built with Hugging Face (ConvFill Talker + Gemma 4 31B reasoner + Phonon 31M TTS + Qwen3-ASR 1.7B STT) is live on HF Spaces.
More from Multimodal
- Creator builds Snorricam video series with Seedance 2.5, sharing prompts — LudovicCreator · 2026-10-27
- Gradium launches free webpage TTS Chrome extension powered by its 40M-parameter Phonon model — mattturck · 2026-10-27
- Vortex of Infinite Recursion prompt goes viral: images that spiral endlessly into themselves — azed_ai · 2026-10-27
- 170+ AI Creators Made One Collective Film, Premiering November 6 — Loo_Atreides · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Midjourney creator showcases --sref style reference outputs: gothic character art — OVolosin82152 · 2026-10-27