Fullduplex Signals
Weekly signals in speech-to-speech and full-duplex voice AI. Latest: 2026-W47, Nov 09 - Nov 15, 2026. Archive: fullduplex.ai/signals
Paper • 2611.12118 • PublishedNote 2026-W47 · A simulated user speaks at set moments during tool execution and spoken result delivery. 60 English scenarios, six domains, 80 tools; five commercial agents tested (GPT Live 1.0, Realtime 1.5, Realtime 2, Grok Voice 3.0, Gemini 3.8 Live, as named in the paper). Under request changes the highest result completion rate is 55%. After an interruption, four of five agents omit at least 23% of the information on average. Interruption latency 1.07-2.53 s. No code or data link found.
An LLM Is All You Need: Data-Efficient Adaptation to Full-Duplex, Word-Level Spoken Dialogue
Paper • 2611.14774 • PublishedNote 2026-W47 · Keeps Qwen3-Omni frozen and trains 97M parameters (adapters and new token embeddings) into a word-level full-duplex thinker on an 80 ms clock, in about 30 GPU-hours. Full-Duplex-Bench v1.0 pause-handling turn-over rate 0.170 on the synthetic subset against 0.584 for the next-best open system; InteractionBench groundedness 53%. The authors report PersonaPlex and Moshi respond faster. Weights on Hugging Face (openmartdev/AwakenFullDuplex-Qwen3Omni, created November 11); code Apache-2.0,
Kezek: Splitting Mono Two-Person Conversations into Per-Speaker Tracks for Full-Duplex Dialogue Data
Paper • 2611.16933 • Published • 2Note 2026-W47 · Splits a mono recording of exactly two people into one track per speaker with the original timing. Against pyannote plus DialogueSidon it keeps more short replies (0.83 vs 0.69 on the Open Yap test set; 0.80 vs 0.71 on CallHome English) and runs at about 230 hours of audio per GPU-hour. For replies more than 10 dB below the other voice it keeps 0.286, where an ideal mask keeps 0.743. English only, offline. Code MIT (github.com/Manifestro/kezek), weights Apache-2.0 (manifestro/kezek-v1
higurashi-ai/Haru-4B
Audio-to-Audio • 5B • Updated • 3.37k • 358Note 2026-W47 · Created November 12, Apache-2.0. Qwen3.5-4B text backbone; every 133 ms (7.5 Hz) it takes one user audio frame and emits one frame of its own audio and text. Audio is a continuous 64-dimension latent per frame from a one-step MeanFlow head, with no discrete audio tokens. Nine languages, about 12 GB GPU memory. The card gives no benchmark numbers and says the model "does not follow instructions the way an assistant does". Self-reported in the release blog: about 65 ms latency on a MacB