The Connect layer led the week with 5 of 16 filings, from LiveKit, Gupshup, Genesys and Quiq.
STT model and option wave beside the OTel break.
livekit-agents 1.8.0 adds AssemblyAI universal-3-6-pro streaming, Deepgram Flux numerals/profanity/redact options, ElevenLabs Scribe v2 secondary_languages, and Google chirp_3 custom prompts.
livekit-agents 1.8.0
$0.10/hr launch price with diarization and timestamps. Microsoft frames it as limited-time through year-end. Treat “fastest/cheapest in the world” as marketing.
Microsoft launched MAI-Transcribe-2, a multilingual ASR model with diarization, timestamps, keyword biasing, and launch pricing of $0.10 per hour of audio.
MAI-Transcribe-2
MAI-Transcribe-2: the highest quality, cheapest transcription at the fastest speed! 10x faster that GPT-Transcribe. Now available on Microsoft Foundry.
https://x.com/MicrosoftAI/status/2095521860184363074
Joint streaming ASR, diarization, and endpointing in 80ms chunks. Vendor: #1 on Artificial Analysis streaming STT as of 1 Sep.
Meta released Muse Voice Transcribe, a streaming ASR model that jointly does transcription, 20+ speaker diarization, and endpointing with multilingual code-switching.
Muse Voice Transcribe
Introducing Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs.
https://x.com/AIatMeta/status/2094839236016976028
New TTS plugin.
Production-focused normalisation and 150 ms p95 TTFA at 48 kHz on 3 September — aimed at the account-number and Hebrew edge cases that break containment, not another demo voice.
Deepdub launched Phantom Z 3.4 Conversational, a multilingual real-time TTS model with 48 kHz audio, stronger English and Hebrew text normalisation, and a vendor-claimed p95 time-to-first-audio of 150 ms.
Phantom Z 3.4 Conversational
Two-tier realtime TTS with prompt tags and LiveKit-native wiring. Flash TTFA (~25ms) is a Coval figure, not a bench we ran.
Inworld launched a two-tier realtime TTS family with natural-language delivery steering, cloning, 200+ language cross-voice support, and on-demand pricing from $25 per million characters.
Realtime TTS-2
Closes the Jul beta on pronunciation of phones, emails, and IBANs before the later latency-focused Gradium drops in September.
Gradium promoted a new default TTS model that passes 81.0% of a 500-sentence hard-case eval across five languages, with Coval TTFA P50 of 216 ms and no caller-side text normalisation.
Gradium TTS (hard-case default)
Breaking observability contract for Agent Insights and third-party GenAI backends across Python and Node.
livekit-agents and @livekit/agents 1.8.0 align spans, metrics, and logs with OpenTelemetry GenAI semantic conventions and move PII filtering in-process; the Node core package ships the same 1.8 line.
livekit-agents / @livekit/agents 1.8.0
CRM glue at the call boundary without Zapier/n8n. Pre-call budget: 1 minute inbound/web, 5 minutes before an outbound dial.
Retell shipped native pre-call and post-call workflows that pull CRM/helpdesk context into voice agents and write outcomes back, without Zapier or n8n.
Retell Workflows
Batch outbound with event-sourced state and a skip webhook, instead of a blob/counter. Lists up to 10,000 contacts.
Vapi rebuilt outbound Campaigns on a durable workflow engine for lists up to 10,000 contacts, with scheduling windows, dynamic variables, and pre-dial skip webhooks.
Vapi Campaigns
Vapi Campaigns is live, rebuilt from the ground up.
https://x.com/Vapi_AI/status/2094832678235209734
Warm transfer on Twilio connector.
livekit-agents 1.8.0 adds Twilio connector warm transfer.
livekit-agents 1.8.0
In several markets the voice agent is a messaging company’s PSTN add-on, not a CCaaS replacement.
Gupshup launched a self-serve console for building and deploying AI voice agents across PSTN and WhatsApp voice alongside WhatsApp, RCS, and SMS, with a choice of STT, TTS, and LLM.
Voice AI Platform
The CCaaS SKU enterprises already bought is getting frontier TTS, STT, and an action model on a published timetable, with ElevenLabs still later this fiscal quarter.
Also Orchestrate
at Xperience, Genesys announced Agentic Virtual Agent upgrades: Scaled Cognition APT-2, end-of-turn detection, Contextual Tuning, and integrations with ElevenLabs and Deepgram.
Agentic Virtual Agent (APT-2 + voice) · vendor ElevenLabs / Deepgram / Scaled Cognition
Genesys is advancing agentic orchestration for CX: context + governance, a smarter Agentic Virtual Agent
https://x.com/Genesys/status/2095182564998709342
This assist product can send the voucher during the call, which is a different job from a whisper bar that only suggests text to the agent.
Quiq extended its AI Assistants to live voice: adaptive next-best-action plus the ability to send customer-facing messages (links, vouchers) on the same thread while the human stays on the line.
Voice Assist
Teams already on Twilio Studio, or on WhatsApp calling, can attach a LiveKit agent without standing up a SIP trunk.
LiveKit launched Connectors that bridge LiveKit rooms to Twilio Programmable Voice and WhatsApp Business Calling without provisioning a SIP trunk, generally available on LiveKit Cloud.
LiveKit Connectors
LiveKit Connectors are generally available today. They bring calls from @twilio Programmable Voice and @WhatsApp Business Calling into a LiveKit room as a regular participant, so your agent can answer there with no SIP trunk to provision and nothing new to host.
https://x.com/livekit/status/2094905369063755830
A meeting-notes vendor running the live call is a different buyer motion from CCaaS or voice-agent infra.
Fireflies launched Voice Agents that conduct recruiting, sales, and support calls from a team knowledge base rather than only taking notes, claiming 40,000+ conversations across 2,100+ organisations in 97 countries since rollout.
Voice Agents
Introducing Fireflies Voice Agents: AI that handles repetitive conversations like screening interviews, sales discovery, customer support, user research
https://x.com/firefliesai/status/2095179947455480113