Local real-time talking-head voice chat lands in audio.cpp 0.9.4 with ~850ms latency on RTX 5090
Acceptable-Cycle4645 · reddit · 2026-11-16
- audio.cpp 0.9.4 adds a real-time avatar voice chat app in its WebUI, combining three models: SoulX-FlashHead for live talking-head generation, OWSM v4 for speech-to-text, and LFM3-Audio for conversational responses and TTS.
- LFM3-Audio runs twice: the S2S path only offers one voice, so its TTS re-synthesizes the reply in a different voice; ASR runs in parallel.
- Performance: 850ms from question to avatar speech on RTX 5090; SoulX needs RTX 4090+; cuDNN is mandatory — without it inference is 2x slower and VRAM usage doubles.
- The double LFM3-Audio call and prototype UI overhead leave room for a native app to cut latency further.
More from Embodied
- Physical AI hits the factory floor: reliability, safety standards and energy per task decide scale — ingliguori · 2026-11-16
- Motorizing ball joints could dramatically simplify humanoid robots — _Stocko_ · 2026-11-16
- Kid attacks occupied Waymo in Miami amid street takeover drama — deanwball · 2026-11-16
- Report: Apple to launch Siri smart glasses in 2027 — nima_owji · 2026-11-16
- PaperILike: Human Interventions as Both Explicit and Implicit Robot Feedback — tomssilver · 2026-11-16
- Gurman: Apple smart glasses prototypes run pared-down visionOS 28, dubbed 'AirPods for your face' — Scobleizer · 2026-11-16