What happened this week
Note: this is a backfill issue. The scheduled task did not run on the Jun 22 publish slot; the cron has since been moved from Mon 09:00 JST to Mon 12:00 JST. W26 itself was a heavy week — five distinct foundational papers plus two substantive platform releases.
Full-duplex — a survey arrives, and Moshi gains a face
A Survey of Full-Duplex Spoken Dialogue Systems (Lu, Wang, Luo) opens the week with the synthesis the field has been needing. More than a dozen systems have claimed to be "full-duplex" in the last year, but the term covers substantially different capabilities. The paper argues this ambiguity is taxonomical — current terminology does not specify where duplex decisions are made, which interaction types are supported, or how a system behaves moment by moment — and introduces three complementary frameworks: an L0-L3 architectural hierarchy, an interaction ontology, and a decision state machine. The cleanest taxonomy of the FD landscape published to date and the natural reference for any new architecture paper from now on.
Moshi-Face (Jiang, Ohashi, Higashinaka) is the more concrete companion. It extends Moshi from audio-only into a full-duplex dialogue model that jointly processes user audio and facial input while simultaneously generating speech and facial motion. A VQ-VAE face codec encodes 3D head meshes into discrete face tokens, fed into the same Moshi-style RQ-Transformer alongside audio. Ohashi appears here again (also W25 author on the Kyutai interactivity alignment paper) — the Moshi research surface is widening fast.
Voice agents — text-only adaptation
CORTIS (Choi, Kim, Kwon) targets the task-oriented voice agent problem: SLMs that map spoken user requests to semantic frames, executable actions, and function calls usually require paired speech-target annotations to learn new tasks. CORTIS is a text-only adaptation framework that fine-tunes the LLM head on text-only data and adapts the speech encoder via an alignment loss, removing the speech-target bottleneck. Practical relevance for anyone deploying SLM-based voice agents into new verticals.
Audio safety — over-refusal joins the gallery
AOR-Bench (Yang, Chun, Lucas) adds a complementary axis to the audio-safety category that has been building since W21. The question: do LALMs over-refuse pseudo-harmful queries — incorrectly rejecting benign requests because they sound harmful in isolation? The audio domain makes this especially hard because speech that appears harmful in isolation may become benign in context. Fits naturally on the new /benchmarks#audio-safety page alongside SpeechJBB (code-switched), AIA (paralinguistic priors), and the LALM jailbreak taxonomy.
Speech-LM analysis
Interleaved Speech Language Models Latently Work In Text (Sternberg, Maimon, Adi) probes how speech-text interleaved SLMs actually represent the two modalities in latent space. Using the logit lens, the paper reveals that these models go through an implicit text-mediated layer even when processing speech-only inputs — a concrete confirmation of the modality-gap-from-the-input-side thesis W20's TextPro-SLM was arguing on different grounds.
Product — LiveKit Turn Detector v1.0 and Pipecat 1.4.0
livekit-agents 1.6.1 introduces LiveKit Turn Detector v1.0 — LiveKit's own state-of-the-art turn detector that leverages both audio and text semantics to decide the optimal moment to respond. The blog framing is that responding too early interrupts the user, responding too late introduces awkward silence — a direct in-house solution to the problem W25's Endpoint Anticipation paper attacked from the academic side. The 1.6.2 follow-up the same day adds AssemblyAI universal-3-5-pro, Gemini 3.1 flash TTS, Soniox stt-rt-v5, and Fish Audio s2.1-pro defaults.
Pipecat 1.4.0 takes a different angle on the same problem. It adds a RealtimeServiceMetadataFrame broadcast at pipeline start by realtime LLM services (OpenAI Realtime, Azure Realtime, Inworld, Grok/xAI, Gemini Live, AWS Nova Sonic, Ultravox) advertising whether the service emits its own user-turn frames — and ships locally-driven-turns examples for each so app developers can choose between server-side and local turn detection. A startup warning log nudges providers whose realtime APIs don't expose ground-truth turn signals. Adds the on_user_turn_message_added event handler to align user-aggregator callbacks across cascade and realtime modes.
What is not here
ESPnet3 (arXiv 2606.21854) is a significant infrastructure release for speech research but sits one layer below the STS / FD digest scope. No in-window dataset drop with a verifiable primary source. Cartesia, Hume, Deepgram Voice Agent, and ElevenLabs Agents did not publish in-window technical changelog items.
Corrections to hello@fullduplex.ai.