hotwirednews Get GrokBot

Week 39

21–27 Sep 20262026-W39 · 79 filings · 62 companies · 27 net-new

The Use layer led the week with 28 of 79 filings, from NatWest Group, Knowtex, EliseAI, UWM and 24 more.

Hearspeech-to-text18

LiveKitPlatformlaunch2026-09-26

LiveKit Agents adds AssemblyAI Universal-3.6 Pro across Agents 1.8.3 and JS 1.9.1

AssemblyAI Universal-3.6 Pro lands on both Python Agents 1.8.3 and JS Agents 1.9.1 the same day; one hear card covers the pair.

Details2 sources

@livekit/agents 1.9.1 and livekit-agents 1.8.3 both register assemblyai/universal-3-6-pro on inference STT; JS adds agent context carryover, and Python surfaces language_confidence on AssemblyAI transcripts.

@livekit/agents 1.9.1 + livekit-agents 1.8.3

Sarvam AIModel providermodel release2026-09-25

Sarvam releases Saaras V4 ASR with five transcript modes across 22 Indian languages

Pipecat wired saaras:v4 in 1.11.0 on 17 September (2026-09-17-pipecat-1-11-0-stt-knobs), eight days before this vendor primary; the WER and language-ID numbers are Sarvam-reported, and the side-by-side samples against Scribe v2, Nova-3, and GPT-4o Transcribe were picked by Sarvam.

Details1 source

Sarvam AI released Saaras V4, a speech recognition model that pairs an audio encoder with an in-house 3B hybrid state-space decoder, served over REST, batch, and streaming WebSocket APIs for Indian-language and English voice apps.

Saaras V4

Daily / PipecatPlatformAPI2026-09-25

Pipecat 1.12.0 defaults XAISTTService to grok-voice-transcribe-2.0

Silent model drift ends; builders get transcribe 2.0 unless they pin 1.0.

Details3 sources

pipecat-ai 1.12.0 sends XAISTTService.model to xAI and defaults to grok-voice-transcribe-2.0; NVIDIA STT settings updates now reapply on the live stream.

pipecat-ai 1.12.0

NKENNEAiModel providermodel release2026-09-25net-new

NKENNEAi launches Swahili speech-to-text and text-to-speech models

Africa-first STT/TTS pair for Swahili (100M+ speakers) from an NSF SBIR Phase II–backed lab — next languages Yoruba, Igbo, Nigerian Pidgin and Somali, with a public demo on 29 Sep rather than GA API docs at capture.

Also Speak

Details3 sources

NKENNEAi, the AI arm of African language platform NKENNE, announced its first speech models — Swahili STT and TTS — joining its translation APIs, with a public live demo set for 29 September 2026.

Swahili STT and TTS

DeepgramModel providerAPI2026-09-25

Deepgram lets Flux STT toggle numerals mid-stream without reconnecting

Stops IVR-style digit capture from forcing a reconnect mid-call on Flux — limited to the listed flux-general languages, so multilingual agents outside that set still need another path.

Details2 sources

Deepgram added a mid-stream Configure.numerals flag on Flux STT so agents can switch digit versus word transcription for PINs, phone numbers, or order IDs without opening a new connection.

Flux STT mid-stream numerals

SpeechmaticsModel providerlaunch2026-09-24

Speechmatics opens early access for on-device speech-to-text

On-device ASR changes the deployment boundary from hosted speech recognition to local inference; the early-access label means availability is not yet general release.

Details1 source · 1 X post

Speechmatics opened early access to a local speech-to-text product that runs without a cloud connection.

Speechmatics On-Device

Run speech-to-text locally. No cloud. No internet connection. Your audio stays on the device.

https://x.com/Speechmatics/status/2103123162112344320
Daily / PipecatPlatformdemo2026-09-23

Pipecat Battleships demo slots NVIDIA Nemotron 3 Diarization into a multi-speaker pipeline

A dated, open repo demo is the clearest hear-layer proof of diarization inside Pipecat at Nemotron 3 launch, composing Smart Turn + PhoneLLM + Jev rather than STT alone.

Also OrchestrateThinkSpeak

Details4 sources · 3 X posts

Kwindla (@kwindla) and Jon Taylor (@JonPTaylor) demo Pipecat Battleships using NVIDIA Nemotron 3 Diarization with Nemotron 3.5 ASR, Smart Turn, PhoneLLM, Typesafe Jev, and Magpie TTS; code at github.com/pipecat-ai/nemo3-battleships.

Nemotron 3 Battleships diarization demo · vendor NVIDIA

NVIDIA released a fast speaker-labeling model today: Nemotron 3 Diarization. Like all of the Nemotron family from @NVIDIA_AI, this is an open weights model, so you can grab the weights from @huggingface and use the model in the cloud or on your local device. Nemotron 3 Diarization is a useful building block for voice agents and realtime AI applications that need to handle speech from multiple people at the same time. It handles overlapping speech well, can identify up to eight speakers, and pairs with any transcription model you're using in your multi-modal pipelines. @jonptaylor built a demo (code below) and recorded a technical deep-dive showing how to use this diarization model together with a speech-to-text model. Jon's code also uses Jev for super-accurate confirmation of critical voice input in a noisy/multi-speaker environment.

https://x.com/kwindla/status/2102796440242413907
https://x.com/kwindla/status/2102796442175988063

Here is @Rekka and I giving Nemo3 Diarization a spin - Pipecat Battleships!‍ ⚓️🦜 Has been a while since I built anything with diarization and this model is really impressive. Also, super easy to slot into a Pipecat pipeline. ➡️ Code: https://github.com/pipecat-ai/nemo3-battleships ➡️ Full 'how it works' video: (attached)

https://x.com/JonPTaylor/status/2102802634226696666
NVIDIAModel providermodel release2026-09-23

NVIDIA opens Nemotron 3 Diarization at VoiceArena #1 (~14.72% DER)

Open streaming diarization that pairs with any ASR — a building block for multi-party agents, with VoiceArena's initial DER lead still subject to the bench's Version 1 completion note.

Details3 sources · 1 X post

NVIDIA released open-weight Nemotron 3 Diarization (~100M params), ranking #1 on VoiceArena's initial Diarization-Bench at 14.72% DER with up to eight speakers in streaming or offline mode.

Nemotron 3 Diarization

When several people talk at once, a transcript can get messy fast. Our new Nemotron 3 Diarization model tracks who spoke when, even when voices overlap. It handles up to eight speakers, has 100M parameters, and is now available on @huggingface 🤗

https://x.com/NVIDIAAI/status/2102775666366435450
KlangModel providermodel release2026-09-23net-new

Klang releases Pianissimo, an open Swedish speech-to-text model built on NVIDIA Parakeet

Open Swedish ASR gets a domestic hosted path (Berget AI processes data inside Sweden and stores nothing) alongside local use; Pianissimo shipped the same day as Klang's SEK 15 million raise and is the first of several open models Klang says it will release.

Details4 sources

Klang released Pianissimo, an open-weight speech-to-text model for Swedish and Swedish dialects that runs locally on an ordinary laptop and is also offered as an API through Berget AI on Swedish infrastructure.

Pianissimo · vendor NVIDIA

KlangModel providerfunding2026-09-23net-new

Klang raises SEK 15 million at a SEK 150 million valuation for in-house real-time speech models

A small insider round (about €1.32 million) aimed at owning the speech model rather than reselling third-party ASR; the open Swedish Pianissimo model released the same day is the first public output.

Details2 sources

Klang, a Helsingborg company that transcribes, summarises, and analyses conversations, raised SEK 15 million from existing owners led by entrepreneur Johan Lenander to fund its own real-time speech and language models and international expansion.

iFLYTEKModel providermodel release2026-09-23net-new

iFLYTEK ships Spark-ASR-2.0 speech recognition model

Chinese open-platform ASR ship beside Alibaba Qwen-Audio 3.1 — vendor-claimed dialect and noisy-scene gains, with only a ~10% cost bump over Spark-ASR-1.0.

Details4 sources

iFLYTEK released Spark-ASR-2.0, a non-autoregressive plus LLM-enhanced autoregressive ASR stack, with API access on the iFLYTEK Open Platform and Input Method rollout from 24 Sep 2026.

Spark-ASR-2.0

ElevenLabsModel providermodel release2026-09-23

ElevenLabs ships Scribe v2 Medical for clinical speech-to-text

Gives builders a clinical-vocab STT SKU on the existing Scribe API without a separate realtime path — ambient charting and intake remain batch, with HIPAA controls gated to Enterprise BAA customers.

Details2 sources · 1 X post

ElevenLabs released Scribe v2 Medical, a batch STT fine-tune of Scribe v2 for clinical audio, claiming a 35% lower word error rate on clinical speech while matching base Scribe v2 on everyday Common Voice samples.

Scribe v2 Medical

We're releasing Scribe v2 Medical, our Speech to Text model fine-tuned for clinical audio. Scribe v2 Medical reduces word error rate (WER) on clinical audio by 35% compared to our base Scribe v2 model, improving recognition of medication names, anatomy, and pathology.

https://x.com/ElevenLabs/status/2102762665995325741
Edge0Model providermodel release2026-09-23

Edge0 open-sources Audio8 ASR Infinite for 24/7 streaming speech recognition

A 30 s rolling KV window with exact RoPE re-basing keeps memory and latency flat for always-on agents — vendor AISHELL-1 CER 1.75 at 480 ms delay versus 16.80 for Voxtral-Mini-4B-Realtime on the same card, while English LibriSpeech still trails Voxtral.

Details5 sources · 1 X post

Edge0 released Audio8 ASR Infinite, a native streaming Chinese and English ASR model with a rolling KV cache for unbounded 24/7 transcription and selectable 80/120/160 ms audio clocks.

Audio8 ASR Infinite

Today we're open-sourcing Audio8 ASR Infinite: Ultra-low latency, unlimited audio, 24/7 transcription, no drift.

https://x.com/SamuelZengML/status/2102726739449332221
ArgmaxPlatformlaunch2026-09-23

Argmax ships Pro SDK 3 with Nemotron 3 Diarization and Qwen3-ASR

Puts the same-day Nemotron 3 Diarization drop on Apple on-device APIs with a pre-ASR diarization path — useful for multi-party mobile agents without a separate clustering stack.

Also Orchestrate

Details2 sources · 1 X post

Argmax released Pro SDK 3 with Day 0 on-device support for NVIDIA Nemotron 3 Diarization (up to eight speakers) and multilingual Qwen3-ASR across 30 languages, plus a Pre-diarized Transcription API.

Argmax Pro SDK 3 · vendor NVIDIA

Introducing @NVIDIAAI Nemotron 3 Diarization on Argmax SDK! > Day 0 support > Real-time diarization with 8 speakers > Frontier accuracy Details in comments.

https://x.com/argmax/status/2102778012622311704
Alibaba / QwenModel providermodel release2026-09-23

Alibaba ships Qwen-Audio-3.1 ASR models with series-wide price cuts

Completes the Qwen-Audio-3.1 stack in the corpus after the 20 Sep Realtime Plus filing — dated ASR GA plus the advertised ASR price floor matter for China-region voice pipelines.

Details4 sources · 1 X post

Alibaba’s Qwen team published the Qwen-Audio-3.1 ASR line (including ASR-Flash and ASR-Next) on Model Studio and announced series price cuts of up to about 95% on ASR.

Qwen-Audio-3.1-ASR (ASR-Flash / ASR-Next)

TwilioPlatformAPI2026-09-22

Twilio announces Real-Time Transcriptions provider failover and Auto mode

Gives CPaaS transcription multi-provider resiliency without custom routing — note the 22 Oct go-live is three weeks after the 22 Sep announce date.

Also Connect

Details1 source

Twilio announced that from 22 October 2026 Real-Time Transcriptions will fail over to a second speech provider at session start when the primary cannot connect, and will add an Auto transcriptionEngine that picks provider per session.

Real-Time Transcriptions failover and Auto

MoondreamModel providermodel release2026-09-22

Moondream releases Parakeet Redux and Parakeet Ultra open-weight speech-to-text models

Redux trades noise robustness for size (9.04% WER in background noise against 6.72% for the original), so CPU deployments gain throughput at a cost on noisy calls; Ultra is the GPU option and beats the base checkpoint on every vendor-reported group.

Details3 sources · 1 X post

Moondream released two CC-BY-4.0 post-trained versions of NVIDIA's parakeet-tdt-0.6b-v3 for 25-language ASR: Redux, a 178 MB ternary model for CPUs and Apple silicon, and Ultra, a full-precision model trained further for GPUs.

Parakeet Redux and Parakeet Ultra · vendor NVIDIA

Today we're releasing Moondream Parakeet Redux and Moondream Parakeet Ultra. Two speech-to-text models based on NVIDIA's Parakeet. 25 languages. They run locally, wickedly fast.

https://x.com/moondreamai/status/2102494472106119401
ModulateModel providerAPI2026-09-21

Modulate opens Velma Accent Detection API for acoustic accent classification

This is a distinct API launch from Modulate’s 28 September funding announcement: it exposes one Velma capability for builders rather than describing the company’s financing or broader audio ensemble.

Details2 sources · 1 X post

Modulate’s Velma Accent Detection API classifies a speaker’s accent from audio and returns a whole-file label plus a per-window time series in one synchronous call.

Velma Accent Detection API

Velma's Accent Detection API classifies a speaker's accent straight from audio, in a single synchronous call.

https://x.com/modulate_ai/status/2102051978159227133

Seevision input1

MOSIModel providerlaunch2026-09-23

MOSS-VL-Realtime opens its browser-to-GPU interaction stack

MOSI pairs MOSS-VL-Realtime with SGLang-Omni for concurrent video calls, yet the real-time API and online Playground it promises have not shipped.

Also Think

Details2 sources · 1 X post

The open pipeline combines browser voice and video calls, screen sharing, interruption, SGLang-Omni inference, and long-term video memory.

MOSS-VL-Realtime

🚨MOSS-VL-Realtime now runs through a complete browser-to-GPU stack @Open_MOSS 💡Realtime Demo: a browser app for live voice/video interaction — video calls, screen sharing, and interruption, no installation needed 🛠️ SGLang-Omni: a specialized inference backend that keeps real-time video calls smooth under concurrent users 🧠 Long-term memory: sliding-window video memory + on-demand recall, so the assistant remembers what you talked about even in long conversations → Bring up the full pipeline with a one-click installer The full model-to-application pipeline is now open source. The third_party directory also includes a FlashAttention-3 module customized for MOSS-VL; mainline @sgl_project support for offline batch inference is already merged into upstream SGLang. 🔥Stay tuned for our real-time API and online Playground without needing to prepare your own GPU! More details below ↓

https://x.com/MosiAI_Official/status/2102699884285690301

Speaktext-to-speech8

LiveKitPlatformlaunch2026-09-26

LiveKit Agents adds Gemini 3.8 TTS across Agents 1.8.3 and JS 1.9.1

Gemini 3.8 TTS + expressive mode ship on both Python 1.8.3 and JS 1.9.1 the same day.

Details2 sources

@livekit/agents 1.9.1 adds gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts with expressive mode and multi-speaker configs; livekit-agents 1.8.3 wires Gemini 3.8 TTS models into expressive mode on the Python path.

@livekit/agents 1.9.1 + livekit-agents 1.8.3

TelnyxPlatformAPI2026-09-25

Telnyx adds Soniox TTS with about 200 voices across 60 languages

Pairs Soniox STT already on Telnyx with Soniox TTS on the same CPaaS path — a same-vendor listen/speak stack rather than a fourth synthesis vendor bolted onto Call Control.

Also Connect

Details1 source

Telnyx Voice AI now offers Soniox streaming TTS as a catalog provider, with about 200 voices that each speak 60 languages and telephony-native 8 kHz PCM output.

Soniox TTS on Telnyx Voice AI · vendor Soniox

Daily / PipecatPlatformlaunch2026-09-25

Pipecat 1.12.0 adds IPA pronunciation transforms for TTS

Drug names and brands get consistent pronunciation without per-provider SSML forks.

Details3 sources

pipecat-ai 1.12.0 adds TTSService.pronunciation_transform_ipa() so one word-to-IPA map works across Cartesia, ElevenLabs, Inworld, and Deepgram Aura-2 formatters.

pipecat-ai 1.12.0

Navana.aiModel providermodel release2026-09-23

Navana.ai opens Bodhi TTS publicly for ten Indian languages with sub-100 ms first audio

Public Indian-language TTS at ₹12/10k chars and sub-100 ms TTFA against Sarvam Bulbul and global multilingual TTS — STT remains invite-only the same day.

Details3 sources

Navana.ai opened public access to Bodhi TTS on its developer platform: more than 50 Indian voices across ten languages at ₹12 per 10,000 characters, with zero-shot cloning and on-prem inference aimed at BFSI call volumes.

Bodhi TTS

GoogleModel providermodel release2026-09-23

Google ships Gemini 3.8 Flash TTS and Flash-Lite TTS with promptable voice design

Puts promptable TTS and 30-second cloning inside the Gemini API the same week as Gemini 3.8 Live — a direct rival stack for ElevenLabs-style voice design, with vendor-claimed Hume #1 scores still subject to Google's own benches.

Details3 sources · 2 X posts

Google released Gemini 3.8 Flash TTS and Flash-Lite TTS in the Gemini API and AI Studio: promptable voice design across 100+ languages, 30-second replication with consent checks, line-by-line direction, and SynthID watermarking.

Gemini 3.8 Flash TTS / Flash-Lite TTS

Can you hear that? Our Gemini Audio family is getting louder 🔊 We're introducing two of our most expressive audio generation models yet from @GoogleDeepMind: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS.

https://x.com/Google/status/2102781485824590115

Create and deploy custom audio with our new text-to-speech models: 🔵 Gemini 3.8 Flash TTS: Design unique voices with distinct accents and characteristics. 🔵 Gemini 3.8 Flash-Lite TTS: Built for efficiency and scale, choose from your created styles or our expansive library.

https://x.com/GoogleDeepMind/status/2102781530867126505
Fish AudioModel providermodel release2026-09-23

Fish Audio previews Drama 3, a TTS model directed in plain language

Fish Audio moves from the inline tags of its open-source S2 (2026-03-09-fish-audio-s2-open-source, W11) to free-text performance direction, and the preview warning that takes vary points at narration and drama work before low-latency agent use.

Details5 sources · 3 X posts

Drama 3 (preview) takes natural-language directions for tone, pacing, and character, can shift delivery mid-sentence, voices multi-speaker scenes in one pass, and can redo a single word.

Drama 3 (preview)

Introducing Drama 3 (preview): the most controllable TTS model ever.

https://x.com/FishAudio/status/2102805658005635294
https://x.com/FishAudio/status/2103311575176347945

Try the preview now in the API with 'drama-3-preview'

https://x.com/FishAudio/status/2107878623126356245
Alibaba / QwenModel providermodel release2026-09-23

Alibaba ships Qwen-Audio-3.1-TTS in Flash and Plus on Model Studio, with instruction and tag control and token billing

Ten weeks after Qwen-Audio-3.0-TTS (2026-07-14-alibaba-qwen-audio-3-0-tts, W29), 3.1 Flash drops per-character billing for $0.23/$1.87 per million tokens in Singapore and is already the default in Model Studio's TTS code samples; with Qwen-Audio-3.1-ASR (2026-09-23-alibaba-qwen-audio-3-1-asr) and Realtime Plus (2026-09-20-alibaba-qwen-audio-3-1-realtime-plus, W38), it gives Model Studio a 3.1 model for each stage of a cascaded pipeline beside the speech-to-speech path.

Details7 sources · 2 X posts

The text-to-speech model in Alibaba's five-model Qwen-Audio-3.1 release is served as qwen-audio-3.1-tts-flash for real-time voice assistants and customer service and qwen-audio-3.1-tts-plus for premium voices and content creation, with multilingual and dialect synthesis, natural-language instructions, and inline emotion tags.

Qwen-Audio-3.1-TTS (qwen-audio-3.1-tts-flash and qwen-audio-3.1-tts-plus)

⚡ Meet Qwen-Audio-3.1! ASR, TTS & Realtime are fully upgraded, joined by two new models: TTS-Next for audio creation and ASR-Next for audio understanding. Five models, one complete audio stack: understanding, generation, interaction & creation.

https://x.com/TONGYI_SpeechAI/status/2104391711166546243

🎉Qwen-Audio-3.1-TTS took #1 on @ArtificialAnlys (Controlled Voice Arena) with an Elo of 1182 https://t.co/j9IMqKyTDd

https://x.com/TONGYI_SpeechAI/status/2108033857627046326
GradiumModel providermodel release2026-09-22

Gradium ships a TTS beta claiming under 50 ms time to first audio

Vendor claim under 50 ms TTFA on the shared `gradium-tts-beta` opt-in path — still unverified here against Coval/Speko numbers, and the July beta brief never advertised that latency.

Details4 sources · 2 X posts

Gradium opened a TTS beta selectable as `gradium-tts-beta` in Studio and the API, stating vendor-measured time to first audio below 50 ms while keeping the same quality bar as its production path.

Gradium TTS beta (sub-50 ms TTFA)

Gradium TTS beta model is out now to define what fast means. Less than 50ms TTFA, same quality. Most models are >100ms TTFA. Try it now: https://t.co/OoQhTtIF0P https://t.co/rXyjKjS4oN

https://x.com/GradiumAI/status/2102438831668551687

check out the benchmarks at @covaldev and @speko_ai

https://x.com/GradiumAI/status/2102439595522572524

Showavatars, video3

GoogleModel providermodel release2026-09-24

Google ships Gemini 3.8 Live with Live Avatar in Gemini Enterprise

Puts embodied S2S video beside the 15 Sep Gemini 3.8 Live API drop — the hyperscaler answer to Meta Muse Realtime Avatar, with enterprise allowlisting and SynthID required before custom faces ship.

Also Think

Details4 sources · 2 X posts

Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise: near-real-time lip-synced video avatars on the native speech-to-speech Live stack, with async tool calling and mid-conversation switching across 97 languages.

Gemini 3.8 Live with Live Avatar

Power your agents: Gemini 3.8 Live with Live Avatar is now generally available in Gemini Enterprise. Key capabilities include: video avatars, fluid dialogue, tool calling, and more. Start building today ↓ https://t.co/MD9Smf29Hq

https://x.com/GoogleCloudTech/status/2103157502988746877

RT @GoogleCloudTech: Power your agents: Gemini 3.8 Live with Live Avatar is now generally available in Gemini Enterprise. Key capabilities…

https://x.com/GoogleDeepMind/status/2103176711479402748
MetaModel providermodel release2026-09-23

Meta ships Muse Realtime Avatar for live video chat with Muse

Extends the prior Muse outbound-calling filing (17 Sep) from telephony into embodied S2S video — same speech-token stream drives voice and lip/body motion, with sub-second serving numbers Meta published for GB200.

Details2 sources

Meta Research published Muse Realtime Avatar, an audio-driven Diffusion Transformer that turns Muse Realtime Voice speech tokens into synchronised portrait video for live Muse conversations.

Muse Realtime Avatar

LemonSliceModel providermodel release2026-09-23

LemonSlice releases Character World Model-1 for controllable live avatars

This is the flagship the 2 October Lite post calls “last week,” and the controllability baseline Lite says it keeps.

Details1 source · 1 X post

CWM-1 is a causal few-step video diffusion transformer that generates a character’s face, body, hands, and scene live during a conversation, with an emotion engine, free in the web app and on the API for Enterprise and Ultra.

CWM-1

Introducing Character World Model-1 the most controllable interactive avatar model in the world Features and how we built it 👇️

https://x.com/LemonSliceAI/status/2102821939441975456

ThinkLLMs, speech-to-speech4

LiveKitPlatformlaunch2026-09-26

LiveKit Agents 1.8.3 adds Azure GPT-Live duplex

Azure brings DuplexModel to builders already on GPT-Live.

Details1 source

livekit-agents 1.8.3 adds GPTLiveModel.with_azure for Azure OpenAI duplex sessions.

livekit-agents 1.8.3

telliPlatformlaunch2026-09-21

telli makes Duo agents the default for new voice agents

Making the split between thinking and speaking the default for every new agent is a bet that two-stage turns give more precise calls; telli names no model vendors, so the components cannot be checked.

Also SpeakOrchestrate

Details1 source

Duo agents separate thinking from speaking and are available to every account as the default for new agents.

Duo agents

NVIDIAModel providerpaper2026-09-21

NVIDIA opens NemotronLabs VoiceChat 11B full-duplex S2S with tool calling

First open full-duplex S2S drop that keeps conversation flowing while tools execute — tool routing leads the cited FDB 3.0 numbers, but argument grounding and end-to-end Pass@1 still trail closed realtime stacks.

Details3 sources

NVIDIA published the NemotronLabs VoiceChat paper and open 11B full-duplex speech-to-speech weights with a parallel tool-calling channel, barge-in, and ~450 ms turn-taking on Full-Duplex-Bench.

NemotronLabs VoiceChat 11B

KyutaiModel providermodel release2026-09-21

Kyutai releases Voice of Reason speech-native math models

Shows RL can lift a speech-native math path from 27.3% to 77.1% GSM8K without a text cascade — still a specialist (TriviaQA speech fell), and the headline numbers use synthesised questions rather than live telephony audio.

Details5 sources · 1 X post

Kyutai released two GLM-4-Voice checkpoints that take spoken math problems and answer in speech without a transcription cascade or a separate text LLM at inference time.

Voice of Reason (GLM-4-Voice of Reason / STITCH)

We're releasing Voice of Reason, a speech-native model that does math out loud. Give it a spoken problem without transcription nor text LLM in the loop, and it reasons and answers in speech. GSM8K goes from 27.3% for GLM-4-Voice to 77.1%. Link in 🧵 https://t.co/kwLdTug8jW

https://x.com/kyutai_labs/status/2102080831242060123

Orchestrateframeworks, evals12

LiveKitPlatformlaunch2026-09-26

LiveKit Agents adds warm transfer and tool-result upgrades across Agents 1.8.3 and JS 1.9.1

Python tool-result hardening and JS warm-transfer/telemetry land the same day; one orchestrate card covers both package trains.

Details2 sources

@livekit/agents 1.9.1 adds TwilioConnectorWarmTransferTask and deeper turn telemetry; livekit-agents 1.8.3 adds reply_required tool results and verify_spelling on GetEmailTask.

@livekit/agents 1.9.1 + livekit-agents 1.8.3

UpfirstPlatformlaunch2026-09-25

Upfirst lets one account run multiple AI receptionists

One account can now separate locations or brands without giving up shared operating knowledge or a unified call log.

Also Use

Details1 source

Upfirst added multiple agents under one account so a business can give separate locations, lines, or businesses their own phone agent while sharing common training.

Multiple agents under one account

Daily / PipecatPlatformlaunch2026-09-25

Pipecat 1.12.0 adds Pipecat Classifiers, Jev evals, and empty-turn recovery

Fast typed judgments replace many bespoke LLM-prompt gates in evals, voicemail, and UI workers.

Details3 sources

pipecat-ai 1.12.0 adds Classifiers (JevClassifier, LLMClassifier), classifier-backed EvalJudge and VoicemailDetector, empty_user_turn recovery, and frame-level interruptible.

pipecat-ai 1.12.0

SoundHound AIPlatformlaunch2026-09-24

SoundHound ships OASYS Edge for on-device agentic voice AI

Moves OASYS from cloud-only agentic CX into OEM edge SKUs — but go-live is still late 2026, so this is a dated architecture launch with CES 2027 demos rather than a fleet-wide ship today.

Also Connect

Details3 sources · 1 X post

SoundHound released OASYS Edge, an embedded OASYS architecture that runs LLM-powered voice agents on vehicle and smart-device hardware with local processing when the network is unavailable.

OASYS Edge

NEWS: SoundHound AI Introduces OASYS Edge, Bringing Fully Embedded Agentic Voice AI to Vehicles and Smart Devices https://t.co/ixzHe5GvW1 https://t.co/9J6iOdgQlV

https://x.com/SoundHound/status/2103110453983662374
MicrosoftPlatformlaunch2026-09-24

Microsoft puts voice agents into Foundry Agent Service public preview

Moves Microsoft from MAI speech SKUs into a governed agent runtime with Teams Phone and Twilio paths — public preview, not partner-only MAI Realtime.

Also Connect

Details2 sources

Microsoft opened public preview of voice agents as a native agent type in Foundry Agent Service, with the same API and SDK as other Foundry agents and deployment to web, Teams, Teams Phone, and Twilio telephony.

Foundry Agent Service voice agents

LiveKitPlatformacquisition2026-09-24

LiveKit acquires Loophole Labs to cut AI agent startup times

The 2026 press-wire primary that earlier sweeps lacked when the deal was X-only — an infra acquihire under the agent runtime, not an Agents SDK feature.

Details1 source

LiveKit announced it acquired Loophole Labs and its eight-person engineering team, planning to put Loophole's virtualization technology into production to target agent startup times under three seconds.

Loophole Labs acquisition

Alibaba / QwenPlatformlaunch2026-09-23

Qwen Audio Agent 2.0 rebuilds its runtime around ACP and A2A backends and ships with a technical report

The same week brought Qwen-Audio-3.1 ASR (2026-09-23-alibaba-qwen-audio-3-1-asr), and 2.0.0 lists Qwen3.8 Omni Flash Realtime (2026-09-20-alibaba-qwen3-8-omni-flash, W38) among nine frontends that include OpenAI and Google models; the 91.04% task-success figure is Alibaba's own 134-case cockpit set, so it does not compare with τ-Voice or other shared benches.

Also Think

Details3 sources

Version 2.0.0 of Alibaba's open-source voice runtime adds a unified client protocol, A2A alongside ACP for backend agents, more voice and video frontends, a desktop conversation panel with remote mobile access, and customer-service, cockpit, and digital-human examples.

Qwen Audio Agent 2.0

RimeModel providerother2026-09-22

Rime open-sources AlphaBench, SupportBench, and ContentBench for task-specific TTS evaluation

Vendor-run and Rime-versus-each-rival only, yet it publishes a loss: narration preference went to both ElevenLabs and Deepgram while support-call preference favoured Rime in all four pairings, so the board argues for task-specific TTS choice rather than a single naturalness score.

Also Speak

Details4 sources

Rime published three open text-to-speech benchmarks that test identifier reading, support-call preference, and narration preference, releasing the corpora, synthesis code, and analysis on GitHub.

AlphaBench, SupportBench, and ContentBench

NVIDIAPlatformlaunch2026-09-22

NVIDIA's Nemotron Voice Agent 2.2.0 adds an OpenAI Realtime-compatible gateway and single-GPU recipes

At v2.2.0, NVIDIA makes a one-GPU speech stack and an OpenAI Realtime-shaped gateway part of the same reference blueprint, but the NVIDIA-only session knobs mean wire compatibility does not make the service vendor-neutral. The release sits in the same week as 2026-09-23-nvidia-nemotron-3-diarization, which supplies a separate diarization model.

Details1 source

The blueprint adds one-GPU speech-and-LLM deployment paths, an OpenAI Realtime-shaped WebSocket, and an agent-building skill.

Nemotron Voice Agent v2.2.0

Ringg AIPlatformlaunch2026-09-21

Ringg launches Spaces for isolated partner and customer agent workspaces

Spaces turns Ringg's multi-channel agent platform into a partner operating layer with isolation and central administration; it clears the conditional product-launch check rather than being filed as an undated admin feature.

Also Use

Details2 sources · 1 X post

Ringg Spaces gives partners and enterprises separate AI-agent workspaces for each client or team while centralising access, costs, and analytics.

Ringg Spaces

Today, we’re launching Ringg Spaces for our partner ecosystem. Create independent sub-workspaces for your customers, with clear separation of data and costs—all managed from one central account. More customers. Less admin. More time building agents that get the work done. https://t.co/VkKi9gH813

https://x.com/ringg_ai/status/2102024450954957137
CekuraPlatformlaunch2026-09-21

Cekura and Ocular open Converse-STT across 15 speech models

Public-clip WER and conversational WER pick different leaders — AssemblyAI Universal 3.5 Pro at 1.93% on Pipecat versus Reson8 at 2.93% on Ocular Hi-Fi — so agent builders choosing STT from a single public board risk the wrong model for real two-party speech.

Also Hear

Details3 sources · 1 X post

Cekura published Converse-STT with Ocular AI, scoring 15 speech-to-text models on word error rate, time to first text, and final-text delay across 1,000 Pipecat clips and Ocular's two-person conversational recordings.

Converse-STT · vendor Ocular

I'm excited to announce that we've launched @cekuraAi's speech-to-text benchmarks: Converse STT, comparing 15 models on transcription accuracy and speed.

https://x.com/tarush_agarwal_/status/2102411594005434521
Applied Brain ResearchModel providerlaunch2026-09-21net-new

Applied Brain Research ships ABR SDK with on-device Niagara ASR and Nith TTS

On-device ASR/TTS at vendor-claimed 115 ms / 147 ms with a five-language GA SDK against cloud-only realtime stacks — still commercial-licence gated, and Cortex-M RTOS targets are promised only before year end.

Also HearSpeak

Details4 sources

Applied Brain Research made the ABR SDK generally available as a single Python-over-C API for streaming speech recognition and synthesis that runs entirely on edge hardware, with no network hop at inference time.

ABR SDK (Niagara ASR + Nith TTS)

Connecttelephony, channels5

Daily / PipecatPlatformAPI2026-09-25

Pipecat 1.12.0 adds MOQTransport client-mode reconnect

MoQ calls survive relay blips instead of ending the session on the first drop.

Details5 sources

pipecat-ai 1.12.0 adds MOQ client-mode redial with backoff, relay_url dialing, transcript seq/epoch dedupe, and ErrorFrame categories on MOQ failures.

pipecat-ai 1.12.0

Delight.aiPlatformlaunch2026-09-25net-new

Delight ships Universal Telephony for voice agents on business numbers

Contact-centre handoff stays a standard phone transfer into Genesys Cloud and Amazon Connect, with Salesforce, Zendesk and ServiceNow case write-back on 25 September — telephony plumbing for agents already on Delight, not a new speech model.

Also Use

Details1 source

Delight.ai opened Universal Telephony so its voice agents answer and place calls on a business's own numbers via Twilio, Telnyx and local carriers, then hand off by phone number into contact-centre queues with CRM case sync.

Universal Telephony

Helo.aiPlatformlaunch2026-09-24net-new

Helo.ai launches Helo Voice for inbound and outbound enterprise calling

India-focused CPaaS voice SKU beside Helo Convo — advertorial wire on 24 Sep after a GFF booth (8–11 Sep), with no named bank or NBFC go-live in the release.

Also Use

Details3 sources

Helo.ai (VivaConnect) opened Helo Voice, a CPaaS voice-agent platform for inbound support and outbound campaigns in Hindi, English and regional languages, with CRM/API task completion, live human handoff, and SIP/WebRTC/PSTN telephony.

Helo Voice

VoisoPlatformlaunch2026-09-23net-new

Voiso launches AI Voice Agents inside its contact-centre queues for inbound and outbound calls

Adds a CCaaS-native voice-agent SKU billed in AI-minutes from $500/mo — useful against bolt-on bot stacks, with the ±600 ms latency claim still vendor-only.

Details2 sources

Voiso made AI Voice Agents available to all customers as first-class agents that answer, qualify, hand off and log phone calls inside the same routing, CDR and reporting stack as human teams.

AI Voice Agents

Useagents in market28

NatWest GroupEnterpriselaunch2026-09-25net-new

NatWest trials voice-to-voice Spending Insights on its own SLM

UK retail-bank voice-native finance trial on proprietary SLMs — RBS-first before group rollout — separate from the text Fraud Triage Agent already live in Cora.

Also Think

Details2 sources

NatWest Group will trial a generative audio-visual Spending Insights tool — first under Royal Bank of Scotland — so customers can discuss spending and budgeting by voice or text with interruptions and topic changes, powered by the bank's proprietary small language model.

Spending Insights

KnowtexPlatformpartnership2026-09-25net-new

Knowtex named prime awardee on VA ambient scribe enterprise contract

Second Tarush-listed ambient vendor on the VA enterprise vehicle the same week as Abridge — competitive federal ambient market, not a single-vendor lock.

Also Hear

Details3 sources

Knowtex said it was named a prime awardee on the VA’s nine-figure ambient scribe enterprise contract, putting its ambient clinical documentation platform on the same federal ordering path as peer awardees.

Ambient scribe × VA enterprise contract

EliseAIPlatformpartnership2026-09-25net-new

EliseAI brings OpenAI GPT-Live-1 into healthcare voice agents

Healthcare vertical putting GPT-Live-1 into production booking and billing calls — complements the 10 Sep API filing rather than replacing it.

Also Think

Details2 sources

EliseAI, an early OpenAI design partner for GPT-Live-1, said it is putting the full-duplex model into healthcare voice workflows for appointment booking, billing questions and EHR-connected tasks.

GPT-Live-1 healthcare voice · vendor OpenAI

UWMEnterpriselaunch2026-09-24net-new

UWM ships ChatUWM mobile app with AI Voice Assistant for brokers

Wholesale brokers get spoken access to live pipeline facts — including three-day lock expiries — from a phone app on 24 September, an operator end-use rather than a vendor platform launch.

Details2 sources

United Wholesale Mortgage replaced its InTouch broker app with ChatUWM, adding a spoken assistant that answers pipeline, lock-expiry, loan-status and guideline questions without typing.

ChatUWM AI Voice Assistant

SurFox AIPlatformlaunch2026-09-24net-new

SurFox AI adds Voice AI that answers, qualifies, and books inbound lead calls

SMB lead tools are pitching the receiving end of agent-placed calls in the same week Google opened Gemini Call for Me on Pixel 11 (2026-09-24-google-gemini-call-for-me); that link is SurFox's own framing, and the 11 October release names no integration or call volumes.

Details2 sources

Miami-based SurFox AI extended its SMS and website-chat lead agent to the phone, so small businesses can have inbound callers qualified and booked into a Google or Outlook calendar while still on the line.

SurFox Voice AI

Sideband StudioPlatformlaunch2026-09-24net-new

Sideband Studio ships hushtype for local Mac dictation with optional cleanup

A local dictation path with optional cleanup gives Mac users a private alternative to hosted voice-typing services; cleanup can send the transcript to the endpoint the user configures, so the offline boundary applies to recognition rather than every feature.

Also Hear

Details5 sources

Sideband Studio’s hushtype lets Mac users hold a shortcut, speak, and paste cleaned text into the active app, with speech recognition running locally and optional LLM cleanup.

hushtype

ServiceTitanEnterpriselaunch2026-09-24

ServiceTitan expands Voice Agent with automatic unsold-estimate recovery

This turns the 2025 inbound voice agent into a post-call revenue-recovery workflow, distinct from the later ServiceTitan-on-Vapi customer story (2026-11-12-servicetitan-vapi-voice-agent).

Also Connect

Details1 source

ServiceTitan’s ST-81 release adds automatic follow-up for unsold estimates to its Voice Agent while also introducing self-serve Automations.

Voice Agent unsold-estimate recovery

PayWithPlayPlatformlaunch2026-09-24net-new

PayWithPlay launches CallGPT voice agents on Nigerian business phone lines

Africa-local PSTN voice agent for SMEs and corporates — same connect+use pattern as Access4 AIX and My Friend, priced and languaged for Nigeria rather than a US CCaaS add-on.

Also Connect

Details3 sources

PayWithPlay launched CallGPT, an AI voice agent that answers existing enterprise phone lines in Nigeria for routine enquiries, bookings, and structured call logging without an app download.

CallGPT

KonectaEnablerlaunch2026-09-24

Konecta launches V2V realtime voice translation for Ley SAC languages

BPO ships agent-in-the-loop voice translation for Spain's Ley SAC deadline rather than a full autonomous voice agent — 60+ languages and no-audio-retention claims are vendor-stated.

Also Connect

Details1 source

Konecta launched V2V, a realtime voice-to-voice translation layer so contact-centre agents keep control of the call while customers hear more than 60 languages, aimed at Spain's Ley SAC co-official language rules.

V2V (Voice to Voice)

GoogleModel providerlaunch2026-09-24

Google Gemini Call for Me places business calls from a Pixel 11

Pixel-side agent calling lands a week after Muse and Instinct Concierge opened consumer outbound dialling (2026-09-17-meta-muse-outbound-calls, 2026-09-16-instinct-concierge-calling), but only for US Pixel 11 subscribers in Phone beta — and the call uses the user's own number rather than a platform trunk.

Also ConnectThink

Details3 sources

US Pixel 11 owners on a paid Gemini plan and the Phone by Google public beta can ask Gemini to dial businesses, navigate menus, wait on hold, and complete errands such as bookings and stock checks while they watch a live transcript.

Gemini Call for Me

Dextr AIPlatformlaunch2026-09-24net-new

Dextr AI exits stealth with Daisy, a hotel voice reservations agent, and a $6.7M seed

Adds a vertical hotel voice-reservation SKU with named PMS integrations and PCI voice payments — file as platform go-live, not funding-only, because Daisy is already taking live booking calls.

Also Connect

Details3 sources

Dextr AI publicly launched its hospitality agent platform, led by Daisy for phone reservations and modifications in 90+ languages, alongside a $6.7 million Elevation Capital–led seed.

Daisy (Voice AI & Reservation Agent)

NiCE CognigyPlatformlaunch2026-09-24

NiCE Cognigy opens India GA with Mumbai-hosted multilingual voice agents

Puts a CCaaS voice-agent stack onshore in ap-south-1 after Deepgram's 15 Sep India endpoint — residency plus local speech for BFSI/CX buyers who would not send calls offshore.

Also Connect

Details2 sources

NiCE made Cognigy generally available in India on a dedicated Mumbai environment (Hyderabad DR), with local STT/TTS tuned for major Indian languages, dialects and code-mixed speech across voice, text and WhatsApp.

NiCE Cognigy India (IN1)

SelaPlatformfunding2026-09-23net-new

Sela raises $21M for voice AI agents that help originate mortgage loans

Sela applies voice agents to a regulated, high-value sales workflow rather than general contact-center support; the release gives a funding total but no independent performance audit.

Also SpeakThink

Details2 sources

Sela announced $21 million across seed and Series A funding for voice AI agents serving mortgage sales teams.

Voice AI mortgage-sales agents

Ringg AIPlatformpartnership2026-09-23

OpenAI case study puts Ringg's agents at 7 million connected calls a month on GPT-5.6

Ringg's Parrot STT (2026-06-05-ringg-parrot-stt) and $15M Series A extension (2026-08-26-ringg-series-a-extension) are already filed; this adds named Indian go-lives with figures (Policybazaar 67%, Groww 72%), all supplied by the vendor and its model provider, and the same page says GPT-4.1 still carries most real-time traffic despite the 90% saving headline.

Also Orchestrate

Details2 sources

OpenAI's customer story describes Ringg's voice, chat, WhatsApp, and web agents for large Indian consumer businesses, routing each task across GPT-4.1 and the GPT-5.6 Luna, Terra, and Sol models.

Ringg agent platform on GPT-5.6 · vendor OpenAI

OpenAIModel providerlaunch2026-09-23

OpenAI adds plugins and ChatGPT Work to ChatGPT Voice on web and mobile

Moves ChatGPT Voice from talk-only into tool-using agent work on the phone — the same day plugins land in Voice — without a new speech model SKU beyond GPT-6 routing already in Work.

Also Think

Details3 sources · 1 X post

OpenAI rolled out ChatGPT Voice with plugins (email, calendar, Slack and other connected apps), GPT-6 Astra/Sol/Luna routing, and Voice inside ChatGPT Work on web and mobile for spoken document and browser tasks.

ChatGPT Voice

We heard you loud and clear. ChatGPT Voice can now: - Use plugins like your email, calendar, and Slack. - Be powered by GPT-6 Astra, Sol, and Luna. - Be used in ChatGPT Work on web and mobile, so you can create docs, decks, sites, and spreadsheets or tackle complex tasks in the browser, just by talking. Rolling out globally today in the latest version of the app.

https://x.com/OpenAI/status/2102808325742322002
My FriendPlatformlaunch2026-09-23net-new

My Friend launches AI phone companion for seniors via ordinary phone calls

Consumer companion on the phone network (connect + use) for an audience that will not install an app — date rests on Chang’s roundup plus a live product site, not a dated company blog.

Also Connect

Details2 sources

My Friend made its AI phone companion available for seniors: a dialable number for 24/7 conversation and medication-style reminders with no smartphone, app, or internet required.

My Friend

Liberty GlobalEnterprisepartnership2026-09-23net-new

Liberty Global signs a three-year Sierra partnership for AI agents across its European telecom brands

Puts another European telco group with ~80m connections onto Sierra beside the SoftBank Japan exclusive (14 Jul 2026) — still a framework with phased go-lives, not a single-brand production metric.

Details2 sources

Liberty Global agreed a three-year framework with Sierra so its operating companies can deploy the same conversational agents on chat, voice and text, with a phased rollout across roughly 80 million fixed and mobile connections already under way.

Sierra AI agents (group partnership) · vendor Sierra

GuavaPlatformmodel release2026-09-23net-new

Guava ships Daytona voice model and publishes Guava Voice Index v1.0

Adds a human-referenced voice-agent scoreboard beside Artificial Analysis and τ-Voice — useful even with the caveat that Guava both runs the index and ranks first on it.

Also Orchestrate

Details3 sources

Guava released Daytona, its proprietary voice-agent model stack, and published Guava Voice Index v1.0 scoring Daytona at 58.88/100 against three peer systems and live human agents on the same calls.

Daytona / Guava Voice Index

Access4Platformlaunch2026-09-23net-new

Access4 launches AIX Voice Assistant for ANZ partners

Regional CCaaS voice-agent GA with published onshore edge latency claims (≤250 ms saved) — file as platform go-live beside Voiso’s same-day AI Voice Agents ship.

Also Connect

Details2 sources

Access4 made AIX Voice Assistant available to Australia and New Zealand partners via SASBOSS, running AI receptionist answering, routing, and post-call transcription on Melbourne and Sydney edge nodes.

AIX Voice Assistant

Yuma AIPlatformlaunch2026-09-22net-new

Yuma AI launches Voice AI that edits Shopify orders on live phone calls

Vertical ecommerce voice SKU that executes Shopify-side actions mid-call, with broader availability gated to 1 October 2026 after a limited beta.

Also Connect

Details3 sources

Yuma AI extended its ecommerce support agent to phone, so the agent can identify callers, pull order history, and apply address changes, cancellations, refunds, and subscription updates while the customer is still on the line.

Yuma Voice AI

TeslaEnterpriselaunch2026-09-22

Tesla brings Grok Connectors to its cars for hands-free inbox, calendar, and app tasks

The car becomes a voice front end for a general agent that acts across a rider's accounts, and the 11 October switch-on put it in every Cybercab robotaxi; neither Tesla nor xAI has published task-success or spend-control figures for in-car use.

Also Think

Details4 sources · 4 X posts

Grok in Tesla vehicles can now use Connectors to manage email, tidy a calendar, and talk through files, chats, and tasks hands-free, with the Grok @bot agent placing food and restaurant orders.

Grok in Tesla: Connectors and Grok @bot · vendor xAI

.@Grok in your Tesla can now do meaningful work for you With Connectors, you can manage your inbox, clean up your calendar, or talk through existing files/chat/tasks – all hands-free

https://x.com/Tesla/status/2102430656349544590

Grok @bot is now integrated into Tesla. Thanks to all the teams working hard on this. It can handle: - DoorDash / OpenTable / Uber Eats for pickup and restaurants on your way home or to work, while discussing what to order with family - Discussing your work and even spinning up a PR - Prepping documents on the way to meet clients And more!

https://x.com/yunta_tsai/status/2102433517796028538

Grok @Bot in your Tesla

https://x.com/Tesla/status/2104957612769063417

Grok @bot is enabled on all Cybercabs and the rest of the Model Y fleet. Please let us know how it helps you and what to improve. Enjoy! 🙂

https://x.com/yunta_tsai/status/2109136211352916249
PASCOMPlatformlaunch2026-09-22net-new

PASCOM ships Pia, a native AI phone assistant for its cloud phone system

DACH UCaaS operators get a native phone-system voice agent with EU hosting and GPT Realtime mini / 1.5 / 2 tiers under a minutes bundle from €0 Start on 22 September — without wiring a separate bot vendor into the PBX.

Also Connect

Details3 sources

Deggendorf UCaaS vendor PASCOM moved Pia from beta to general availability for all customers, with a choice of GPT Realtime models, EU data-centre hosting, and bundled-minute tiers that start at a free Start plan.

Pia (PASCOM Intelligence Agent)

NumaPlatformlaunch2026-09-22net-new

Numa launches Operator, an AI receptionist with Precision Routing for dealerships

Dealership-specific voice SKU that routes by need rather than menu presses — vendor claims 95% correct-destination routing across a 1,300+ store install base, not a general CCaaS agent.

Also Connect

Details3 sources

Numa opened Operator as the inbound front door of its dealership stack, so the agent answers every line, hears why the caller rang, and transfers them to the right person or department with that context intact.

Operator (Precision Routing)

AbridgePlatformpartnership2026-09-22net-new

Abridge wins VA ambient AI enterprise contract path

Largest US public-sector ambient documentation path to date — Abridge already operational on both VA EHR stacks before enterprise ordering opens.

Also Hear

Details2 sources

Abridge was selected to provide ambient AI under the VA’s multiple-award enterprise contract (five-year ceiling $775.72M across vendors), already live on VistA/CPRS and the Federal EHR at more than 75 medical centres from a prior pilot.

Ambient AI × VA enterprise contract

407 ETREnterprisepartnership2026-09-22net-new

407 ETR moves contact centre to Amazon Connect with Accenture NeuraFlash

Named toll-road operator live on Connect with AI self-service on 22 September — an Accenture Edge mid-market go-live, not a generic SI practice announcement.

Also Connect

Details1 source

Ontario toll operator 407 ETR modernised its contact centre onto Amazon Connect with NeuraFlash (Accenture Edge), adding AI-powered self-service and real-time agent support while migrating more than 250 employees.

Amazon Connect contact centre (AI self-service) · vendor Amazon Connect · delivered by Accenture (NeuraFlash / Accenture Edge)

SiaEnablerpartnership2026-09-21net-new

Sia partners with ElevenLabs to package voice-agent hackathons and production builds

ElevenLabs gains a certified European consulting channel for production voice agents; with no client named on 21 September, Sia stays the filed company under the enabler rule rather than an invented operator.

Details5 sources

Sia, an international management consulting group, announced a partnership with ElevenLabs that sells clients a Voice AI Hackathon for use-case prototyping and a Voice Agent Build & Scale offer for production agents.

Voice AI Hackathon and Voice Agent Build & Scale · vendor ElevenLabs

PrestoPlatformpartnership2026-09-21

Presto joins Toast Partner Ecosystem to pipe Voice AI orders into Toast POS

Ties Presto’s drive-thru Voice AI to Toast’s POS install base a week before the 28 September Remus-affiliate raise named the same Toast partnership.

Also Connect

Details1 source

Presto Phoenix joined the Toast Partner Ecosystem so shared QSR customers can send Presto Voice drive-thru orders into Toast POS and kitchen workflows.

Presto Voice × Toast Partner Ecosystem · vendor Toast

InTouchNowPlatformfunding2026-09-21net-new

InTouchNow raises £2.3M seed led by Ada Ventures for NHS GP voice agents

UK primary care enters the patient-access voice bucket of the Cekura healthcare map (28 September) with a funded operator-side vendor; the 150-practice footprint comes from investor and press copy, and no NHS outcome data has been published.

Details4 sources

InTouchNow, which runs AI voice reception for UK general practices, raised £2.3 million in seed funding led by Ada Ventures with Exceptional Ventures and Kadmos Capital.

AI voice reception for GP practices