LiveKit Agents 1.4.0 adds Azure STT TrueText post-processing
STT-only surface for 1.4.0.
Details1 source
livekit-agents 1.4.0 adds a TrueText post-processing option on Azure STT.
livekit-agents 1.4.0
The Hear and Speak layers led the week, with 4 filings each.
STT-only surface for 1.4.0.
livekit-agents 1.4.0 adds a TrueText post-processing option on Azure STT.
livekit-agents 1.4.0
Streaming sibling to the 29 January v4 Async cut — semantic endpointing and manual finalisation target agent barge-in failures that silence-only VAD still causes.
Soniox released stt-rt-v4 for low-latency streaming transcription across 60+ languages, with millisecond finality, semantic endpointing, manual finalization, and simultaneous streaming speech translation.
Soniox v4 Real-Time (stt-rt-v4)
Open-weight Realtime STT under Apache 2.0 at claimed sub-200ms delay gives voice-agent builders a self-host path alongside a $0.003/min batch API — WER claims are Mistral's benches.
Mistral AI released Voxtral Transcribe 2 as Voxtral Mini Transcribe V2 for batch jobs and Voxtral Realtime for live transcription, with diarization, context biasing, word-level timestamps, and Apache 2.0 open weights for Realtime.
Voxtral Transcribe 2
Contact-centre builders get a streamed STT-plus-translation WebSocket on 2 February 2026 — months before DeepL's broader Voice-to-Voice suite in April.
Also Connect
DeepL opened the DeepL Voice API to API Pro customers so applications can stream audio and receive a source-language transcript plus translations into up to five target languages in real time.
DeepL Voice API
TTS providers clustered into one speak row.
livekit-agents 1.4.0 adds a Camb.ai TTS plugin among speak-path provider work.
livekit-agents 1.4.0
India-first TTS with 35+ voices and streaming for agents — Josh Talks telephony preference is third-party; full-band still trails ElevenLabs v3 alpha on the same study.
Sarvam AI released Bulbul V3, a production text-to-speech model with 35+ professional voices across 11 Indian languages, low-latency streaming, consent-based voice cloning, and code-switching for Hinglish and other mixed speech.
Bulbul V3
The $11 billion mark gives the voice-model and agent company a concrete financing signal; the announcement supplies no split between speech models, ElevenAgents, or other business lines.
Also Use
ElevenLabs announced a $500 million funding round at an $11 billion valuation.
ElevenLabs funding round
We raised $500M at an $11B valuation to transform how people interact with technology. https://t.co/bF7K6mYOLZ
https://x.com/ElevenLabs/status/2019048384011854087
Zero-retention removes a deployment blocker for regulated TTS workloads; Inworld explicitly extends the setting to cloned voices and names three compliance regimes, but gives no independent audit detail in the launch post.
Also Use
Inworld made zero-data retention available for enterprise customers using Inworld TTS, including customers in regulated industries.
Zero-data retention for Inworld TTS
Zero-data retention is now available for Inworld TTS For enterprise customers in healthcare, finance, and other regulated industries, you can now ensure no customer text inputs or audio outputs are stored after processing. Here's how it works: When enabled, all text inputs are processed in memory to generate audio, then both the text and audio are immediately redacted from our systems. This applies to everything, including requests using cloned voices. We're also HIPAA, SOC 2 Type 2, and GDPR compliant. Whether you're handling protected health information or sensitive financial data, you get the security guarantees needed for production deployment at scale. Questions about enterprise security compliance? Reach out → https://t.co/BY31XyGmKo
https://x.com/inworld/status/2018738552641843331
Avatar plugins filed as one show cluster. Class-2: folded quiet patch(es) 1.4.5 into this row.
livekit-agents 1.4.0 adds an Avatario avatar plugin and customisable Bithuman GPU avatar endpoint handling.
livekit-agents 1.4.0
The round records fresh capital for an avatar-and-voice platform, while the post leaves the investor and valuation details unspecified.
Also Speak
Synthesia announced a $200M Series E funding round for its AI video and avatar platform.
Series E funding
We planned a surprise internal reveal for our $200M Series E (and failed). Confetti, speech, champagne on ice — everything was ready for the big moment. 20 minutes before it kicked off, we spotted a cake at reception. $200M written on it. In massive letters. It had been sitting there for hours. 😅 Surprise ruined. Cake still tasted great though.
https://x.com/synthesiaIO/status/2018327903071510689
LLM/Responses cluster separate from orchestration APIs.
livekit-agents 1.4.0 adds xAI Responses LLM and Azure OpenAI Responses integrations.
livekit-agents 1.4.0
Major orchestration API surface for the 1.4 line across Python and Node Agents. Class-2: folded quiet patch(es) 1.4.2, 1.4.5 into this row.
livekit-agents and @livekit/agents 1.4.0 add LLMStream.collect() for LLM use outside AgentSession, stable tool IDs, commit_user_turn for manual realtime turns, and AgentConfigUpdate with initial judges; the Node package lands on the same 1.4 feature line.
livekit-agents / @livekit/agents 1.4.0
Voice finally shares the same Coaching and Playbooks path as chat and email from early February — one reasoning layer instead of per-channel stacks.
Also UseConnect
Ada began rolling out a Unified Reasoning Engine that replaces channel-specific stacks with one shared intelligence layer across Voice, Messaging, and Email, bringing Coaching and Playbooks onto voice for the first time.
Unified Reasoning Engine
Shifts public LLM comparison toward the multi-turn, tool-heavy profile voice agents actually need — and shows accuracy leaders still miss the latency budget that production cascades require.
Daily released aiewf-eval, an open benchmark that scores text-mode and speech-to-speech LLMs on tool calling, instruction following, knowledge grounding, and latency over ~30-turn voice-agent sessions.
aiewf-eval LLM voice-agent benchmark
Native EHR ambient scribe from Epic itself — raises the bar for third-party ambient vendors that sell into Epic shops.
Also Hear
Epic released AI Charting, a built-in Art feature that listens during visits, drafts the clinician note, suggests orders from the conversation, and accepts voice commands to reshape note structure.
AI Charting (Art)
The suite puts voice agents inside payor workflows such as benefits questions, care-gap outreach, and appointment scheduling rather than limiting them to generic call deflection.
Also ConnectOrchestrate
Infinitus launched an agentic AI suite that handles payor member communications across inbound and outbound voice and text interactions.
Agentic AI Member Services Suite
Thanks to @FierceHealth for covering the launch of our Agentic AI Member Services Suite for payors: https://t.co/c2oXvQ9EbL
https://x.com/InfinitusAI/status/2022018512332407066
The 99% figure is the bank's own log-based judgement on about 3,000 FAQ calls before launch, not a live-traffic rate; the same line later took SBI Hyper Deposit questions (10 July) and gained operator handoff in October (2026-10-15-sbi-shinsei-bank-ai-window-operator-handoff, W42).
The bank's dedicated line for Bright 60, its programme for older customers, now answers in natural conversation around the clock instead of through a menu-driven voicebot.
Bright 60 AI window (Recho AI Voice Agent) · vendor Recho