The Orchestrate layer led the week with 4 of 15 filings, from Lua, LiveKit, Cloudflare and Daily / Pipecat.
Edge STT at 2.7–13 MB with 37 languages targets offline/privacy devices — MAC and SRAM figures are vendor lab specs.
Sensory launched an on-device speech-to-text engine optimised for TensorFlow Lite Micro and NPUs including Arm Ethos-U55, with 37-language support and two footprints: a 2.7 MB domain-specific model and a 13 MB general-purpose model.
On-device STT engine
Preview TTS with inline audio tags across 70+ languages and SynthID on every clip — Elo 1,211 is Artificial Analysis via Google's post, treat as third-party but vendor-framed.
Google introduced Gemini 3.1 Flash TTS in preview on the Gemini API, Google AI Studio, and Vertex AI, adding natural-language audio tags for style, pace, and delivery across 70+ languages with SynthID watermarking on all audio.
Gemini 3.1 Flash TTS
Voxtral TTS and smoother NVIDIA Magpie turns expand open and hyperscaler speak options at GA.
pipecat-ai 1.0.0 adds MistralTTSService for Voxtral TTS and enhances NvidiaTTSService with cross-sentence stitching within an LLM turn.
pipecat-ai 1.0.0
Dated LiveKit Agents latency write-up for the pre-CWM stack. The later CWM-1 page also cites 471 ms, but that figure is interrupt frame time, not this Flash TTFT protocol.
LemonSlice 2.1 Flash is a video diffusion-transformer avatar model and inference stack that LemonSlice reports at 471 ms average time-to-first-byte, and 2.04 s average end-to-end when paired with third-party STT, LLM, and TTS.
LemonSlice 2.1 Flash
We did it! We built the fastest interactive avatar model Introducing LemonSlice-2.1 𝘍𝘭𝘢𝘴𝘩 ⚡ Here’s how we did it using @modal and @livekit 👇 (note: it was not easy)
https://x.com/LemonSliceAI/status/2045203651635843378
New avatar vendor for Agents. Class-2: folded quiet patch(es) 1.5.2 into this row.
livekit-agents 1.5.3 adds a Runway Characters avatar plugin.
livekit-agents 1.5.3
Presence events prevent talking to an avatar that has not joined yet.
pipecat-ai 1.0.0 adds on_avatar_connected and on_avatar_disconnected on the LemonSlice transport when the avatar joins or leaves the Daily room.
pipecat-ai 1.0.0
LLM provider cluster for 1.5.3. Class-2: folded quiet patch(es) 1.5.7, 1.5.14 into this row.
livekit-agents 1.5.3 adds a Cerebras LLM plugin and xAI Grok LLM support on the inference path.
livekit-agents 1.5.3
Realtime Inworld and persistent Responses sockets become default think-layer paths at 1.0.
pipecat-ai 1.0.0 adds InworldRealtimeLLMService (cascade STT/LLM/TTS with semantic VAD) and makes WebSocket OpenAIResponsesLLMService the Responses API default.
pipecat-ai 1.0.0
Multi-agent supervisor with voice support on the same Agent OS that later added Browser/WhatsApp/Phone/Meetings transports — orchestration first, channel second.
Also Use
Lua launched Spaces: supervisor agents that reason over a fleet of specialised agents, delegate in one conversation, and synthesise a single reply across web, WhatsApp, Slack, Teams, and voice.
Spaces
Realtime handoff reuse is the orchestration headline.
livekit-agents 1.5.3 reuses realtime sessions across agent handoffs when supported, and adds beta InstructionParts plus ToolSearchToolset and ToolProxyToolset.
livekit-agents 1.5.3
Puts voice on the same Durable Object agent as text and tools without a separate framework — experimental, with Workers AI defaults and a Twilio adapter caveat on mulaw audio.
Also HearSpeakConnect
Cloudflare released an experimental @cloudflare/voice package for the Agents SDK so the same Durable Object agent can take real-time voice over its existing WebSocket, with built-in Workers AI Flux/Nova 3 STT and Aura TTS plus a Twilio phone adapter.
@cloudflare/voice
1.0 locks the builder-facing service and frame contract the rest of 2026 builds on.
pipecat-ai 1.0.0 stabilises the Settings-pattern service APIs, async function calls with cancel_on_interruption=False, LLMMessagesTransformFrame, and removes deprecated pre-1.0 surfaces.
pipecat-ai 1.0.0
Consumer/carrier distribution for Kotoba plus a named East-Asia S2S language pair beyond the May API alpha and August Agentforce JP filings.
Also ThinkSpeak
Kotoba integrated its realtime interpretation app into SoftBank's Daredemo AI on 17 April 2026 and officially released a Koto-powered Japanese–Korean speech-to-speech model.
SoftBank Daredemo AI + Japanese–Korean S2S
Speech-to-speech translation suite with staggered GA/early-access SKUs and 40+ languages — Slator 96% preference is DeepL-commissioned.
Also ThinkConnect
DeepL launched DeepL Voice-to-Voice for live spoken translation across meetings, mobile/web conversations, group QR sessions, and a Voice-to-Voice API, expanding DeepL Voice to over forty languages including all 24 official EU languages.
DeepL Voice-to-Voice
Named national QSR operator go-live path — corporate test into select franchisee expansion — for Presto Voice rather than an unnamed pilot.
American Dairy Queen Corporation partnered with Presto on drive-thru Voice AI after corporate-store tests, expanding the pilot to selected franchisees across the US.
Drive-thru Voice AI (Presto) · vendor Presto