Gradium CEO: every voice agent today is a half-duplex walkie talkie
AI Engineer · youtube · 2026-09-15
A talk by Gradium co-founder and CEO Neil Zeghidour argues that every real-time voice model shipping today is half duplex — either listening or speaking — while real phone calls involve up to 20% overlapping speech. That gap is the case for full-duplex voice models.
Key points:
- Traces the arc from a 2011 phone assistant mapping transcripts to app actions, through open-ended voice chat with zero agency, to cascaded voice agents that call tools but inherit text-pipeline latency and flattened emotion
- Live demo: back-channeling ("mm hmm", "yeah") makes current models stop dead, exposing broken turn-taking
- Technical core: 8 words ≈ 3 seconds = 72,000 timesteps at 24kHz; attention cost grows with the square of sequence length, so raw audio can't fit in an LM. Neural codecs compress audio into learnable tokens; full duplex requires modeling two token streams at once
- Candid tradeoff: every naturalness gain costs intelligence under a fixed weight budget. Two paths forward: scale speech-to-speech models, or split interface from intelligence
More from Multimodal
- Midjourney dual --sref combo produces medieval woodcut cat knight art — sergeantsref · 2026-11-02
- Fluffy AI-generated character designs: fox girl, alien bug and cat photographer — OVolosin82152 · 2026-11-02
- Fable 5.5 reportedly one-shots animation in pure code as Anthropic's coding bet pays off — Hesamation · 2026-11-02
- FlowHMR gets 5x faster, demoing an LLM agent pipeline from video to photoreal mocap — camenduru · 2026-11-02
- Retro 1973 music video "Rock is Dead" by TECHHALLA showcases startlingly realistic AI generation — Mother-Pear-2392 · 2026-11-02
- Weekend splatting with COLMAP and Lichtfeld — janusch_patas · 2026-11-02