The Orchestrate layer led the week with 5 of 8 filings, from Daily / Pipecat, LiveKit, xAI, Vercel and 1 more.
Together becomes a one-vendor STT/TTS option beside Deepgram and Cartesia pairs.
pipecat-ai 1.5.0 adds TogetherSTTService and a mode streaming parameter on AssemblyAISTTService for U3 Pro latency profiles.
pipecat-ai 1.5.0
TTFA makes speak-layer latency comparable across providers in the same observer stream.
pipecat-ai 1.5.0 adds TogetherTTSService, per-sentence synthesis and zero-shot prompts on NvidiaTTSService, and Time To First Audio metrics on TTS services.
pipecat-ai 1.5.0
Flow-based bots share the same release train and import path as the rest of Pipecat.
pipecat-ai 1.5.0 brings Pipecat Flows into the core package so conversation graphs install without a separate flows distribution.
pipecat-ai 1.5.0
LiveKit's own latency-first LLM SKU for Agents — 192 ms TTFT vendor claim versus multi-hundred-millisecond hosted peers on the same voice-shaped prompts.
Also Think
LiveKit Inference began serving Gemma 4 31B with voice-shaped serving (long system prompts, large tool catalogs, low queueing) billed on the same LiveKit API key as Agents.
Gemma 4 31B on LiveKit Inference
Packages Grok Voice S2S with telephony and tools under one meter — a platform SKU ahead of Think Fast 2.0 (29 Jul), not another raw model card.
Also ThinkConnectUse
xAI launched Voice Agent Builder in beta: a no-code console that turns a plain-language call flow, documents, tools, and guardrails into a production phone agent on Grok Voice speech-to-speech.
Voice Agent Builder
Puts OpenAI-class realtime voice behind the same Gateway metering Next.js teams already use for text — beta on 29 Jun, before OpenAI's 6 Jul gpt-realtime-2.1 API bump.
Also HearSpeakThink
Vercel put realtime voice agents, text-to-speech, and speech-to-text on AI Gateway in beta through AI SDK 7, with the same observability, spend controls, and bring-your-own-key path as its other modalities.
AI Gateway realtime voice
Moves Retell from no-code build into continuous ops: failed-call → simulation → reviewable patch, with human approval on every live-agent edit.
Retell announced Conductor, a review-first AI copilot that proposes edits inside a production voice-agent workflow, generates simulation tests from failed calls, and ships nothing until a human accepts each change.
Conductor
Browser and mobile clients can dial DTMF without a telephony-only serializer path.
pipecat-ai 1.5.0 adds base_url on TwilioFrameSerializer and a first-class RTVI dtmf client message for keypad input.
pipecat-ai 1.5.0