Open-source voice pipeline adds Smart Turn end-of-turn gate before LLM calls
ivan_digital · reddit · 2026-09-04
The maintainer of speech-core, a C++17 voice runtime, released v0.0.14 with an optional end-of-turn gate based on Pipecat/Daily's Smart Turn v3.2 model, preventing a thinking pause from finalizing a transcript and invoking the LLM prematurely.
How it works:
- VAD detects a candidate pause
- Smart Turn examines up to the last 8 seconds of 16 kHz audio
- At threshold 0.5, the pipeline completes or continues the turn
- Speech resuming after a vetoed pause stays in the same turn
- A 2-second max-silence cap prevents indefinite holds
Implementation details:
- Gate lives in the shared C++ state machine; the C API accepts alternate classifiers via vtable
- speech-swift: 17 MB Core ML model, 3.5 ms per window on Apple Silicon, decisions identical to the fp32 upstream on 1,000 test clips (both 92.9%)
- speech-android: 11.1 MB int8 ONNX model, CPU-only, runs once per confirmed pause
- int8 export scored 93.0% on the same clips
- Eager STT now respects the classifier's veto; flush, forced splitting, classifier failure, and silence cap all handled explicitly
Fully open source, no paid feature gates.
More from coding & agent
- Software World: A Simulated GitHub Ecosystem Run by Agents for Multi-Agent Research — PangWeiKoh · 2026-09-04
- Developer Building 'Path of Codex', an ARPG Made with OpenAI's Codex — Dimillian · 2026-09-04
- GPT-6 demo shows fully autonomous 3D modeling in Blender, shipping scenes to UE5 — petewoodbridge · 2026-09-04
- Gemini CLI vs Crush vs Chrome DevTools MCP: three tools that aren't interchangeable — Different_Regret_146 · 2026-09-04
- OpenClaw's revamped onboarding lets users set up phone and Mac gateways easily — steipete · 2026-09-04
- Instinct launches interface-free personal agent living in iMessage and WhatsApp — Scobleizer · 2026-09-04