LiveKit Agents 1.5.15 adds Cartesia ink-2 STT
Cartesia STT model addition. Class-2: folded quiet patch(es) 1.5.17 into this row.
Details2 sources
livekit-agents 1.5.15 adds Cartesia ink-2 STT.
livekit-agents 1.5.15
The Orchestrate layer led the week with 5 of 20 filings, from Daily / Pipecat, Infinitus, Parloa and ElevenLabs.
Cartesia STT model addition. Class-2: folded quiet patch(es) 1.5.17 into this row.
livekit-agents 1.5.15 adds Cartesia ink-2 STT.
livekit-agents 1.5.15
Cartesia's turn-aware ASR becomes selectable beside Deepgram Flux and AssemblyAI U3 Pro.
pipecat-ai 1.3.0 adds CartesiaTurnsSTTService for Cartesia Streaming ASR with turn signals, plus max_endpoint_delay_ms on SonioxSTTService.
pipecat-ai 1.3.0
Voice input puts Salesforce's agentic work surface behind a spoken interface, extending the platform's voice path beyond customer-service calls and into employee workflows.
Also UseOrchestrate
Salesforce launched Slackbot voice input so people can speak to the AI teammate and use the resulting conversation to act across Salesforce and Slack workflows.
Slackbot voice input
Keeps Rime users on current model names after Arcana defaults in earlier releases.
pipecat-ai 1.3.0 adds the Rime coda model to RimeTTSService and RimeHttpTTSService.
pipecat-ai 1.3.0
On-prem deployment is aimed at buyers whose legal teams will not sign a US vendor's DPA for patient or payment data, a different wedge from latency-led cloud TTS APIs. The January open checkpoint (2026-01-11-kugelaudio-0-open, W02) came first.
The Berlin YC Spring 2026 company opened its real-time text-to-speech model and API for 30+ languages and dialects, with voice cloning and IPA support, hosted or deployed on-prem.
KugelAudio TTS
KugelAudio is launching today! Excited to share our first frontier voice model with the world: real‑time multilingual TTS.
https://x.com/KugelAudio/status/2060121416230699428
Performance-conditioned multilingual speech — not flat TTS from a transcript — is the gap studios and creators still pay humans to close; API still sales-gated at launch.
ElevenLabs launched Dubbing v2, which conditions on the original speaker's performance to carry tone, pacing, and emotion across 90+ languages with sync-aware translation, available in ElevenCreative and ElevenProductions with API access still forthcoming.
Dubbing v2
NineNineSix later based Akylai (2026-10-07-nineninesix-akylai) on MOSS-TTS-v1.5, a documented downstream fine-tune for Kyrgyz and Kazakh.
The update retains the 1.0 capabilities while adding inline pauses such as [pause 3.2s], 31-language synthesis, and more stable zero-shot cloning.
MOSS-TTS-v1.5
Introducing MOSS-TTS-v1.5. Built on MOSS-TTS-1.0, it keeps the full 1.0 capability set and brings more controllable speech generation to multilingual TTS. Now with inline pause control like [pause 3.2s], stronger 31-language synthesis, more stable zero-shot voice cloning, and improved prosody for long-form speech. A more flexible model for professional voice creation. Try it on Hugging Face.
https://x.com/MosiAI_Official/status/2059311099216793721
Diffusion LLM option for builders comparing cascade latency versus autoregressive peers.
pipecat-ai 1.3.0 adds InceptionLLMService for Inception's Mercury 2 diffusion reasoning model.
pipecat-ai 1.3.0
First developer-facing Kotoba surface in the corpus; pairs with the June seed and August Agentforce Japan filing.
Also HearSpeak
Kotoba Technologies released an alpha API and Python SDK exposing speech-to-speech translation plus streaming STT and TTS, aimed at builders evaluating East Asian voice workloads.
Kotoba API & SDK Alpha
Every existing user has to rebuild the local index with refresh --force --reset-index, and the 55% latency gain is the maintainers' own measurement.
Pipecat Context Hub 0.1.0 upgrades its vector store from ChromaDB 0.6 to 1.5.x, whose on-disk format is not backward-compatible with existing local indexes.
Pipecat Context Hub 0.1.0
Subagents and UI workers become a core runtime pattern rather than parallel pipelines glued by hand.
pipecat-ai 1.3.0 makes pipelines multi-agent compatible by default and adds the pipecat.workers framework including UIWorker for observing and driving client web UIs.
pipecat-ai 1.3.0
The new layer targets the dropped context between outreach calls, human agents, and records; its healthcare focus is explicit, but the launch publishes no deployment or outcome figure.
Also Use
Workflows lets healthcare organisations coordinate AI agents, human agents, APIs, copilots, and third-party systems across a patient's or member's journey.
Workflows
NEWS: Today we're announcing Workflows, the next evolution of the Infinitus platform. Learn more: https://t.co/4WylKAptja
https://x.com/InfinitusAI/status/2059998877739004386
Voice-to-chat continuity is the headline claim, but the blog describes the shared foundation for text and other modalities as something to come, and publishes no latency number. The related infrastructure page lists 140+ languages and direct SIP connectivity, with no date.
Also ConnectSpeak
The Conversational Platform puts orchestration, runtime infrastructure, and real-time interaction handling for telephony and web experiences on one foundation, with audio streaming aimed at lower latency and better barge-in handling.
Conversational Platform
Splits the stack so ElevenLabs owns the speech loop and builders keep their own LLM — the missing middle between raw TTS/STT APIs and a fully hosted ElevenAgents runtime.
Also HearSpeak
ElevenLabs released Speech Engine, a WebSocket path that handles STT, turn-taking, TTS, and browser playback while the developer's server owns the LLM logic and streams response text — an alternative to fully hosted ElevenAgents.
Speech Engine
Vonage WebRTC joins Daily, LiveKit, and SmallWebRTC as a first-party transport.
pipecat-ai 1.3.0 adds VonageVideoConnectorTransport for realtime Vonage WebRTC sessions.
pipecat-ai 1.3.0
The GA date turns Salesforce's native voice contact-center proposition from an announced product into a purchasable platform surface, while the source publishes no independent resolution or handle-time measurement.
Also UseOrchestrate
Agentforce Contact Center is generally available as a native system for connecting voice, digital channels, CRM data, and AI agents.
Agentforce Contact Center
OPBX treats AI agents as routing destinations next to extensions and ring groups, so any SIP or WebSocket voice agent can sit behind a self-hosted phone system without a hosted CCaaS; the release names no deployments.
Also Orchestrate
OPBX is an open-source, multi-tenant business phone system on the Cloudonix CPaaS that sends calls to multiple AI voice agents, with real-time load balancing between providers.
OPBX
Spoken complaints now route straight into a portal that took 15.2 lakh grievances in the first half of 2026, though DARPG has published no chatbot usage; it is the first central-government voice service on BHASHINI speech in this corpus, after the open VoicERA stack (2026-02-18-bhashini-voicera, W08), and neither release links the two.
Citizens describe a complaint aloud in their own language, and the CPGRAMS chatbot asks follow-up questions, picks the ministry, department, and category, and files the grievance with the right officer.
Samadhan Didi (CPGRAMS voice chatbot) · vendor BHASHINI
the chatbot enables citizens to navigate their grievance journey through simple voice or text interactions in 22 Indian languages
https://x.com/_BHASHINI/status/2060718371474403668
Hon'ble MoS (PP) Dr. Jitendra Singh will launch the CPGRAMS AI Chatbot tomorrow, 30 May 2026 at Kartavya Bhawan, New Delhi. DARPG, in collaboration with Bhashini, has developed the chatbot "SAMADHAN DIDI" that understands citizen inputs, analyzes grievances, and automatically identifies the relevant Ministry, Department, and grievance category, making the grievance filing process more seamless
https://x.com/DARPG_GoI/status/2060319289580888439
Cart, recall, loan, deposit, and collections journeys make this a concrete vertical expansion across voice and digital channels; Talkdesk gives no independent outcome data for the launch.
Also Connect
Talkdesk's Customer Experience Automation platform now offers outbound AI agents for retail and financial services across voice and digital channels.
Proactive AI Agents for Retail and Financial Services
We’re launching AI agent use cases! Talkdesk’s AI Agents for Retail manage 🛒Cart abandonment 🛒Large-scale recalls Talkdesk’s AI Agents for Financial Services 💰Drive loan and deposit growth 💰Improve collections outcomes Learn more: https://t.co/I8Hpzgo4vw #Talkdesk #CX
https://x.com/Talkdesk/status/2059624445082239232
Searching while speaking is Sesame's answer to the latency-versus-accuracy trade-off in voice agents, comparable in aim (not design) to the split conversation and reasoning layers Sierra later announced (2026-10-06-sierra-tandem-voice); the app grew from 39 countries here to about 100 by the October release (2026-10-06-sesame-voice-assistants-agents).
Also Think
Sesame released an iOS app preview of its conversational voice agents, adding the characters Simone and Charlie beside Maya and Miles, with search cards, notes, texting, Deep Dives, per-character memory, and an Incognito mode, free during the rollout with a possible waitlist.
Sesame app (iOS preview)
A collection of personal agents, crafted for everyday conversation. Preview available now on iOS.
https://x.com/sesame/status/2059698748394262689