Open-source 32M model gives voice assistants human-like "mhm" backchannels in 5ms on CPU
solyarisoftware · x · 2026-11-02
KoljaB released backchannel-cue-32m on Hugging Face (MIT, 129MB ONNX): a tiny model that makes voice assistants interject human-like backchannels ("mhm", "oh no", "right") instead of staying silent until you finish.
- Input is just the user's words plus counters—no audio, no LLM needed.
- Outputs a gate probability for whether to respond, 61 reaction classes, text mood, and likely ask-back wording.
- 5ms on short text, 12ms on 40 words, single CPU thread, callable at every pause.
- Practical tips: call 0.45s into a pause on live transcripts; play pre-rendered audio rather than generating on the fly; hand off to the assistant once the user's turn is clearly complete.
More from Models
- Claim: Opus 5.5 recreates OpenAI's new promo entirely in code with Three.js, GLSL and an SDF heroine — FinanceYF5 · 2026-11-02
- HalachaBench sparks debate: is Talmudic Q&A testing knowledge or reasoning? — teortaxesTex · 2026-11-02
- Leak: Anthropic's unreleased Claude Mythos 5.1 surfaces; Fable 5.5 launch expected this week — haider1 · 2026-11-02
- Community prediction of Opus 5.5 launching in October called 'surprisingly accurate' — dejavucoder · 2026-11-02
- WhatsApp officially congratulates Kimi: 'you made history' — crystalsssup · 2026-11-02
- Remember when everyone called GPT-Astra AGI? Nobody uses it after Opus 5.5 — ns123abc · 2026-11-02