The Use layer led the week with 8 of 33 filings, from Daily / Pipecat, Meta, Betterbot, Assort Health and 4 more.
Puts a hyperscaler STT refresh into the same price band as Transcribe 1.0, with a named Loom deployment and a timed 1.0 sunset — peers will pressure-test the vendor benches.
xAI shipped Grok Voice Transcribe 2.0, a batch and streaming STT model it says is twice as accurate as Transcribe 1.0 at the same $0.10/hr batch and $0.20/hr streaming price.
Grok Voice Transcribe 2.0
A dedicated agent STT surface from a long-running ASR vendor, with both Pipecat and LiveKit plugins available on day one and a public latency–accuracy claim that peers will pressure-test.
Speechmatics released Agent STT, a speech-to-text API for production voice agents, powered by its new Linden 1 model with sub-350 ms finalisation, 55+ languages, live diarisation, and conversational events over `/v2/agent`.
Agent STT (Linden 1)
That’s why we built Agent STT.
https://x.com/Speechmatics/status/2100607821846921567
STT latency/accuracy and vocabulary controls catch up across the major providers.
pipecat-ai 1.11.0 adds OpenAIRealtimeSTT delay, OpenAISTT keywords, Whisper initial_prompt, Sarvam saaras:v4, AWS Transcribe partial_results_stability, and more AssemblyAI language_codes.
pipecat-ai 1.11.0
Splits pharma entity accuracy from general medical STT — a measurable KRR claim competitors in clinical voice will be asked to match.
Deepgram released Nova-3 Pharma, a speech-to-text model trained for drug names and pharmaceutical vocabulary, available for batch and streaming on hosted and self-hosted deployments via model=nova-3-pharma.
Nova-3 Pharma
Introducing Nova-3 Pharma, the first speech-to-text model purpose-built for the pharmaceutical industry.
https://x.com/DeepgramAI/status/2100614929573093694
Bundles transcript-plus-rewrite into one billable call for push-to-talk UIs — a different request shape from streaming agent STT already in the corpus.
AssemblyAI released the Dictation API, a sync endpoint that returns a verbatim transcript and an LLM-cleaned finished text from one short clip on Universal-3.5 Pro, priced at $0.62 per audio hour all-in.
Dictation API
Dear developers, Forms are broken.
https://x.com/AssemblyAI/status/2100710329282101434
Blurt: push-to-talk dictation on the AssemblyAI Dictation API.
https://x.com/AssemblyAI/status/2100971811584516172
ElevenLabs multilingual realtime STT.
livekit-agents 1.8.2 adds secondary languages and language detection for ElevenLabs realtime STT, and migrates Speechmatics to Agent STT.
livekit-agents 1.8.2
Adds a managed onshore option after EU and Australia — relevant for Indian BFSI voice workloads that previously faced self-host or cross-border trade-offs.
Also Speak
Deepgram opened a generally available India regional endpoint at api.in.deepgram.com in AWS ap-south-2 (Hyderabad), with in-country storage and inference for STT, TTS, Voice Agent, and text intelligence at Global/EU/Australia pricing.
India endpoint (api.in.deepgram.com)
The Deepgram India endpoint is now generally available.
https://x.com/DeepgramAI/status/2099894045023641666
Telephony EQ effects and spoken math matter for contact-centre and tutoring bots.
pipecat-ai 1.11.0 adds Azure effect (eq_car, eq_telecomhp8k), Smallest math_notation, and Soniox client_reference_id on TTS settings.
pipecat-ai 1.11.0
Moves brand-voice localisation from one-voice-per-locale cloning to accent adds on a single ID — useful for agents that already run Sonic-3.6.
Cartesia launched Multilingual Voices so one cloned or library voice can speak multiple languages and accents natively while keeping the same voice ID, demonstrated with SF bakery Baklavastory.
Multilingual Voices
Introducing Multilingual Voices:
https://x.com/cartesia/status/2100630777407091096
https://x.com/krandiash/status/2100630035208323305
Rime Coda streaming TTS.
livekit-agents 1.8.2 adds Rime Coda streaming support.
livekit-agents 1.8.2
After Phoenix-4.5, CVI gets relationship continuity: post-call consolidation plus inspectable/editable stores, with Pinned Memories for handoff context.
Also Orchestrate
Memories gives each PAL–person pair a Profile and Timeline that update after calls, so returning conversations continue with prior goals, events, and preferences without re-explaining.
Memories (for PALs)
New avatar vendor; telemetry-heavy lines left unfiled. Class-2: folded quiet patch(es) 1.8.1 into this row.
livekit-agents 1.8.2 adds a Synthesia interactive-avatar plugin.
livekit-agents 1.8.2
Vendor-claimed 2.3 s LAAL and speaker-attributed cloned speech on a 60/29-language realtime API — a direct peer for Gemini Live Translate-style interpretation workloads.
Alibaba Cloud and the Qwen team released Qwen3.8-LiveTranslate, a WebSocket realtime interpretation model that understands 60 languages and speaks 29, with average lagging cut from 2.8 s to 2.3 s versus the prior generation.
Qwen3.8-LiveTranslate (qwen3.8-livetranslate-flash-realtime)
Makes 3.1 the documented default for Bailian realtime voice agents while leaving Flash as the cost-sensitive option; migration is a model ID and voice swap on the existing 3.0 protocol.
Alibaba Cloud Model Studio documents qwen-audio-3.1-realtime-plus as the recommended end-to-end speech-to-speech model for voice assistants and customer-service conversations, available in Singapore and China (Beijing).
Qwen-Audio 3.1 Realtime Plus
Gemini 3.8 Live becomes selectable the same week Pipecat UI lands for builders.
pipecat-ai 1.11.0 adds gemini-3.8-live and gemini-3.8-live-extended-thinking on GeminiLiveLLMService with NON_BLOCKING tools and multi-turn_complete hold.
pipecat-ai 1.11.0
Shanghai lab posts a duplex S2S API family that leads AA conversational dynamics while still trailing peers on first-audio latency — the quality/latency trade is the story.
Also HearSpeakOrchestrate
StepFun released StepAudio 3 Realtime, a full-duplex audio-language model with Think-While-Speaking and a Voice Agent, reporting 98.9 overall on Artificial Analysis Full-Duplex Bench alongside ASR Max, TTS, Gen, and Music siblings.
StepAudio 3 Realtime
A hyperscaler S2S drop with async tools + visual grounding, same day both Pipecat and LiveKit wire it — the competitive reference for GPT-Live-1-class agent stacks, with Artificial Analysis and τ-Voice numbers that need independent checking.
On 15 Sep Google released Gemini 3.8 Live and 3.8 Live Extended Thinking for native speech-to-speech on the Gemini Live API and AI Studio, with async tool calls, visual grounding, and mid-conversation switching across 97 languages.
Gemini 3.8 Live and 3.8 Live Extended Thinking
Moves meeting translation from shared synthetic voices to per-speaker identity — a different product axis from agent STT/TTS already tracked here.
DeepL shipped new Voice models that preserve each speaker's voice, tone, rhythm, and expression during live multilingual translation, initially across 14 languages, plus a desktop app for Zoom, Teams, and Meet.
DeepL Voice (voice preservation)
Introducing voice preservation in DeepL Voice.
https://x.com/DeepLcom/status/2099831527404154912
Funds a single-model face-to-face duplex bet against cascaded avatar stacks — research preview still ahead, but NVIDIA joined the round.
Also ShowSpeak
Seattle’s Nuance Labs raised a $50 million Series A led by Lightspeed, with Accel, South Park Commons, NVIDIA, and Define Ventures, to build one full-duplex audiovisual model for face-to-face conversation.
Nuance Labs foundation model
Release eval packs become compact multi-case files with marker-aware assertions.
pipecat-ai 1.11.0 lets scenario files hold many cases under scenarios:, adds llm_marker expectations, function_call eval judges, and manifest name/concurrency/data fields.
pipecat-ai 1.11.0
Separates identity cloning from controllability. Same model can win motivational casting and lose meditation; Fish's voice-design swings are the clearest example in the board.
Hume published a Voice Controllability Leaderboard scoring 17 TTS models on voice design, instruction-following, inline tags, and role fit with blind human raters across 13 languages.
Voice Controllability Leaderboard
A result from the Voice Controllability leaderboard we published today: Fish's voice design model won nearly every head to head for motivational speaker voices and lost nearly every one for meditation voices. Same model, same prompt format, and the job you're casting for flipped
https://x.com/hume_ai/status/2100584836864024785
Separates brief human judgment from escalation — a concrete HITL pattern for branded voice/chat agents that otherwise double-pay for transfer.
SoundHound launched Human Assisted Resolution (HAR), an opt-in OASYS feature that lets an AI agent ask a colleague a specific question in real time and finish the conversation without a full handoff.
Human Assisted Resolution (HAR)
Today we're launching Human Assisted Resolution (HAR).
https://x.com/SoundHound/status/2100253157674582423
Turns concurrency limits from hard rejects into a held queue — material for telephony agents that spike above seat count during campaigns.
Also Connect
ElevenLabs shipped concurrency call queueing for ElevenAgents: callers over the agent limit hear hold audio and receive queue_status events until admitted or timed out, with APIs to upload custom hold clips.
ElevenAgents call queueing
LiveKit bots can send avatar or UI frames without a second Daily video path.
pipecat-ai 1.11.0 lets LiveKitTransport publish a camera track from RGB/RGBA/BGRA/ARGB frames with codec and max bitrate controls.
pipecat-ai 1.11.0
Cross-CCaaS duplex handoffs without dropping the call — Fall 2026 availability, not GA yet, but the primary wires the Connect↔Agentforce path builders have been stitching by hand.
Also UseOrchestrate
Salesforce and AWS announced Agent2Agent support so Agentforce Voice and Amazon Connect Customer can exchange real-time bidirectional audio over WebSockets for live multi-agent voice handoffs, slated for Fall 2026.
Agentforce Voice × Amazon Connect Customer (A2A) · vendor Amazon Connect
The default builder-facing client is now the Pipecat UI console product.
pipecat-ai 1.11.0 makes pipecat init React clients use Pipecat UI (ui.pipecat.ai) instead of voice-ui-kit, with console connect, transcript, metrics, and live events.
pipecat-ai 1.11.0
Consumer agent stacks now dial merchants, not just chat — product parity with Instinct Concierge (2026-09-16-instinct-concierge-calling), filed from TechCrunch plus the Muse product lead's X post rather than a Meta blog.
Also ThinkConnect
Meta expanded the Muse beta so the agent can place outbound phone calls to US businesses, rolling first to users who had already asked Muse for calling.
Muse outbound business calling
we just expanded the @Muse beta for outbound calls to US businesses, prioritizing folks who'd asked their Muse to let us know they wanted it first :) if you're in, give it a shot and tell us what you think. phone calling was one of our top requests and your feedback helped make it happen! 📞
https://x.com/wailord/status/2100342273854894533
Multifamily operators get phone in the same agentic context as chat and SMS on 17 September, with Entrata/Yardi/RealPage sync — a vertical peer to EliseAI-class housing voice rather than a greenfield CCaaS.
Also Connect
Betterbot opened a new Voice AI lane and a Leasing Dashboard CRM so one agentic stack answers, qualifies, and routes multifamily prospect and resident calls alongside chat, SMS, and email.
Voice AI and Leasing Dashboard
Moves vertical voice agents past intake into write-back scheduling and referrals on a major ambulatory EHR — the bottleneck after the phone is answered.
Assort Health announced expanded Platinum API Tier access so its AI agents can read appointment slots and create, cancel, or reschedule appointments directly in NextGen Enterprise EHR, plus referrals, notes, and payment workflows.
NextGen Enterprise EHR direct integration (expanded) · vendor NextGen Healthcare
Consumer text agents that already run email and trusted-network assistant-to-assistant contact now dial merchants for tasks with no web form — same week Meta Muse opened US business calling (2026-09-17-meta-muse-outbound-calls).
Also Connect
Early-access Concierge has Instinct dial businesses for high-touch errands such as offline restaurant bookings, dentist cancellation lists, and cable-bill disputes.
Instinct Concierge
Introducing Instinct Concierge – a white glove service meant to handle high-touch cases, such as making phone calls, high-end service booking, and more.
https://x.com/noahrshinn/status/2100262985491231101
Packages ElevenAgents as an SMB receptionist SKU (calendar + after-hours) against dedicated receptionist vendors, not only API buyers; HIPAA stays on the broader platform.
Reception by ElevenAgents answers inbound calls, books appointments from a website scan, and goes live with a phone number in minutes; plans start at $29/month with 70+ languages.
Reception (Reception.ai by ElevenAgents)
A consumer voice tutor sold to parents at human-tutor prices — the bet is productive struggle via live voice + whiteboard, not another homework chatbot; outcomes evidence still pending beyond engagement metrics.
Aristotle opened nationwide access to its voice-first AI tutor for students roughly ages 13–18 after a closed beta of 1,000+ students and 1,500+ tutoring hours, alongside a $5 million seed led by True Ventures.
Aristotle voice tutor
Moves a voice-first patient-access vendor into EHR document workflows, so outbound referral calls sit on the same agent stack as inbound phones rather than a bolted-on fax tool.
Hello Patient bought Converse Health so its patient-conversation agents can also do referral and fax intake, chart-driven follow-up, authorisation paperwork, and records work inside outpatient EHRs.
Converse Health acquisition (back-office agents)