turn-taking

Controllable Transcription. Verbatim ( every, filler, pause, stutter, vocal sound) , or intended ( what the speaker meant to say, optimized for readability) with word-level timestamps.

Turn detection for full-duplex dialogue communication

A real-time software for turn-taking, backchannel, and head-nodding prediction

[EMNLP 2026 Main] Speaking While Listening: a survey and empirical audit of full-duplex spoken dialogue systems — L0–L3 architectural hierarchy, T×I×R interaction ontology, and a five-state decision machine, with a curated list of models, datasets, and benchmarks.

Python package based on Gibson's framework (2003) for turn-taking in group conversation analysis.

Local-first VAD, barge-in, and turn-taking primitives for interruptible voice agents.

EgoSpeak: Learning When to Speak for Egocentric Conversational Agents in the Wild

🏀💬 Advanced User Interfaces course project developed as a 1st year student of MSc in CS Eng. at Politecnico di Milano

Full-duplex conversational turn-taking in C++17 — no ASR, no LLM, no neural VAD. Decides when you have finished speaking from prosody alone, and hears you interrupt it through its own speaker echo.

Production C/C++ TurnSense runtime using canonical GGUF and upstream ggml

Open middleware protocol to make voice AI agents sound human — ambient audio, turn-taking, spoken register. Built on Pipecat.

Research preview: timing-aware turn-taking for spoken language models with external Time Adapters and action-token control.

Voice-agent evaluation that tells you which layer broke. Counterfactual blame attribution over a typed call trace, 60 addressable failure modes, and a harness with its own eval.

Predict the silence as a distribution, then decide when to speak. A model-agnostic turn-taking timing layer for spoken dialogue: calibrated silence-survival forecasting + optimal stopping under an asymmetric cost.

Synthetic full-duplex dialogue generation with labels for free, audited by independent turn-taking instruments (VAP, codec control), including the negative results

Conversation vocale en flux continu, sans détecteur de parole : ASR → LLM → TTS, micro-tours à horloge fixe, c'est le modèle qui décide quand parler.

🔄 Adaptive semantic endpointing, four-state conversational machine, and sub-millisecond Barge-In controller for Java voice agents.

Production voice agent stack on a real phone number: real-time speech-to-speech (scribe_realtime ASR, turn_v3, flash TTS) over Twilio PSTN 8 kHz + WebSocket, 16 webhook tools, pinned LLM failover, latency SLOs, eval-gated releases (61-case suite in CI, graded by code), live observability, cost governance. A phone companion persona, with receipts.

MOAS (Machine Operation Alignment System)