Hearspeech-to-text16
SquadStack opens Arth V2 streaming speech recognition for Indian phone calls
An Indian outbound voice-agent platform now publishes its in-house streaming ASR plus an open Hugging Face call set against Sarvam, Deepgram, Google, and OpenAI — still no standalone USD STT SKU on 11 October.
Also Orchestrate
Details3 sources · 1 X post
SquadStack made Arth V2 public as the default STT inside its Humanoid Voice AI Agents, with a vendor-run 67 ms finalize-to-response and 8.85% semantic WER on 863 Hindi-English sales calls.
Arth V2
Introducing Arth V2, SquadStack's in-house streaming speech model, built for India's phone calls 🧵
At 67 ms, it's the fastest model that outpaces every engine and beats Sarvam, Deepgram, Google and OpenAI on accuracy across 863 real Indian sales calls.
https://x.com/apurv_agrawal/status/2109531471715463470
HaloPlatformlaunch2026-10-11net-new
Halo 2.0 adds hold-to-talk dictation and agent status to the Mac notch
Hold-to-talk dictation inside a $9.99 notch utility sits next to dedicated dictation clients such as Featherweight (2026-10-07-featherweight-dictation-launch); Apple on-device recognition is the default, and Groq Whisper runs only if the user adds their own key.
Also Use
Details2 sources
Halo's Mac Dynamic Island update types spoken text at the cursor with on-device Apple speech recognition or optional Groq Whisper, and shows Claude Code and Codex status in the notch.
Halo 2.0
Cactus Whistle adds live streaming and long-file transcription on device
On-device ASR that can stay open on a mic, rather than 30-second clips, is what edge voice agents need next to Needle tool calls — latency figures are Cactus's own device samples, not an independent bench.
Details4 sources · 1 X post
Cactus Compute's Whistle ASR now streams live transcription through needle.stream and transcribes audio longer than the prior 30-second one-shot cap, shipping in cactus-needle 3.2.0 on the same CPU engine as Needle.
Whistle streaming and long-file transcription
VocaScript launches browser transcription for hours-long recordings with speaker labels and quality repair
Long-form browser STT with repair tooling sits beside desktop dictation apps such as Featherweight (2026-10-07-featherweight-dictation-launch) and Willow Knowledge (2026-10-06-willow-knowledge); VocaScript targets multi-hour recordings and URL/extension capture rather than push-to-talk into the cursor, and still hides its model stack.
Details3 sources · 1 X post
VocaScript transcribes long audio and video without forcing splits, returning timestamped, speaker-labelled text in 100+ languages, with a quality report that can retranscribe bad stretches and merge speakers by voice.
VocaScript
TelnyxPlatformlaunch2026-10-09
Telnyx AI Assistants add transcription fallback chains and a second-model language booster on the same call
STT failover and dual-model transcription move from custom pipeline code into an assistant setting, with Telnyx claiming English word errors down from 5.7% to 4.6% using Deepgram Flux plus Nova-3; those are internal tests, not a public bench. The Arabic-English merge rule lands in the same week as other Arabic STT work such as TII Falcon-ASR (2026-10-06-tii-falcon-asr), a comparison only.
Also Orchestrate
Details1 source
Assistants can list up to three backup speech-to-text models that take over when the active one fails, or run a second model beside the main one and choose the better transcript per turn, lock onto the stronger engine, or merge Arabic and English words.
AI Assistants transcription fallback and language booster
Simple AI's Tango model scores caller frustration, confusion, and happiness from call audio
Hume relaunched its Expression Measurement API with new voice models on 6 October (2026-10-06-hume-expression-measurement-api), while Simple AI keeps sentiment inside its own agents rather than selling it as a separate API.
Also Orchestrate
Details1 source
Simple AI added real-time sentiment to Tango, its audio-native conversational awareness model, so its sales and support voice agents rate how callers feel from tone rather than transcripts.
Tango sentiment
LiveKit Agents 1.8.6 adds a Nabrah Arabic STT plugin
Arabic-first builders on LiveKit get a Saudi-hosted STT path with native end-of-turn signalling alongside existing AssemblyAI and Deepgram plugins, without waiting for a framework rewrite; Nabrah's own API launched earlier in 2026 and is new to this corpus as a LiveKit integration.
Details3 sources
livekit-agents 1.8.6 ships livekit-plugins-nabrah, a streaming Arabic and English speech-to-text plugin with client-side end-of-turn detection for Nabrah's recogniser.
livekit-agents 1.8.6 · vendor Nabrah
ElevenLabs adds synthetic voice detection to ElevenAgents and joins Personal Agent Protocol
ElevenAgents now sits with Sierra Fleming (2026-10-06-sierra-curie-fleming-models) and Liberate AI Intercept (2026-09-28-liberate-ai-intercept) on the receiving side of Muse-style outbound calling, though ElevenLabs gives no accuracy figures and enterprise FDE access is the only path until the promised broader October rollout.
Also Orchestrate
Details4 sources · 3 X posts
ElevenAgents can classify a caller as human or AI-generated from the opening seconds of live audio and apply routing rules, while ElevenLabs joins Meta and Sierra's Personal Agent Protocol working group as a design partner.
ElevenAgents synthetic voice detection
A growing share of calls to businesses aren't from humans. They're AI agents acting on behalf of individuals, other companies, and sometimes bad actors.
To help businesses respond, we're releasing synthetic voice detection in ElevenAgents, and joining the emerging Personal Agent Protocol working group to define how agents interact.
https://x.com/ElevenLabs/status/2108666775441461579
We're also launching synthetic voice detection to analyze the caller's speech in the opening seconds of a call, determine if it's human or AI generated, and route accordingly.
This means people get an experience built for people, agents can get direct, bounded interactions, and bad actors can be stopped.
https://x.com/ElevenLabs/status/2108667158792470838
Synthetic voice detection is available now to enterprise customers through our forward deployed team, with broader access later this month and we'll share progress on PAP as the work evolves.
Learn more: https://elevenlabs.io/blog/synthetic-voice-detection-and-personal-agent-protocol
https://x.com/ElevenLabs/status/2108667258293846296
Weird Machines' Off the Record reshapes a Mac user's voice so meeting transcription engines lose the words
Its hiring plan is sold to interviewers, which sets it against candidate-side copilots such as ParakeetAI's Call Assistant (2026-10-07-parakeetai-2-0, filed a day earlier); the claimed defeat of AssemblyAI Universal-3.6 Pro (2026-09-29-assemblyai-universal-3-6-pro-realtime, W40) and Scribe v2 Realtime rests on the maker's own unpublished tests.
Details2 sources
Off the Record is a macOS menu-bar app that adds targeted noise to the user's outgoing microphone audio on the device, so people on Zoom, Google Meet, or Teams calls still follow them while AI notetakers and hidden interview copilots get little usable text.
Off the Record
KrispPlatformAPI2026-10-07
Krisp roughly doubles Voice Isolation calls per CPU server with a one-flag FP16 mode
Self-hosted voice isolation is a CPU cost on every call before the STT, so 76 to 137 calls on one 16-vCPU Graviton3 box roughly halves that line for fleets that can leave Ice Lake; it serves the VI 2.5 models (2026-08-12-krisp-voice-isolation-2-5), and the no-quality-loss claim is Krisp's own.
Details4 sources · 1 X post
With the enableLowPrecisionExecution flag set, Krisp's VIVA SDK runs its Voice Isolation models in FP16 on AWS Graviton3 and Intel Sapphire Rapids CPUs for teams self-hosting noise and background-voice removal in front of voice agents.
Krisp Voice Isolation FP16 execution (VIVA SDK)
Featherweight Dictation launches a 4.6 MB push-to-talk dictation app for Mac and Windows
The tiny client depends on cloud transcription with no offline mode, the opposite trade-off from on-device engines such as Fermion's Phonon-2 Core ML build (2026-10-06-fermion-phonon-2-coreml); every size and speed comparison is Featherweight's own, and its two published prices disagree.
Also Use
Details3 sources
Chennai-based Featherweight released a hold-to-talk dictation app that sends speech to its servers and types cleaned, app-formatted text at the cursor in any Mac or Windows app.
Featherweight Dictation
WillowPlatformlaunch2026-10-06
Willow Knowledge imports writing style and context from ChatGPT, Claude, or Gemini into its dictation app
Context import runs on top of the Frontier dictation models Willow released on 7 July (2026-07-07-willow-frontier-pro-mini, W28) and reaches only macOS and Windows at launch, leaving the iPhone keyboard out; the 190 ms latency comparison posted a day earlier is Willow's own test.
Also Use
Details2 sources · 1 X post
Willow added Knowledge to its macOS and Windows dictation app: users paste a Willow prompt into the AI they already use, bring the reply back, and Scribe applies those preferences and personal details when it writes from speech.
Willow Knowledge
Your AI already knows you. Now Willow can too.
Launching Willow Knowledge today.
Bring your writing style and context from ChatGPT, Claude, or any AI you use into Scribe.
Y’all asked for this. It’s here :)
Available for everyone on macOS and Windows.
We’re also on @ProductHunt. Come check it out and tell us what you think: https://www.producthunt.com/products/willow-voice/launches/willow-knowledge
https://x.com/WillowVoiceAI/status/2107516366492156056
TII ships Falcon-ASR, a 1.6B speech recogniser for Emirati and Standard Arabic
Arabic dialect transcription gains a sovereign-lab entrant that TII scores ahead of Cohere's July 2026 Arabic model on the open Arabic leaderboard, though with only a 60-second demo Space at launch, and no weights or API, builders cannot yet wire it into phone agents.
Details4 sources · 1 X post
The Technology Innovation Institute released Falcon-ASR, a 1.6-billion-parameter transcription model for Emirati Arabic, Modern Standard Arabic, English, French, Spanish, and Portuguese with word-level timestamps, launched alongside the Falcon-Emirati LLM and Falcon-OCR-Arabic.
Falcon-ASR
Falcon-Emirati understands what Emirati means. 84.83% on Alyah, ahead of every Arabic model we compared.
Falcon-Emirati: https://chat.falconllm.tii.ae/?model=Falcon-Emirati-7B
Falcon-ASR: https://huggingface.co/spaces/tiiuae/falcon-asr-demo
Falcon-OCR-Arabic: https://ocr.arabic.aidrc.tii.ae/
Learn More at TII: https://www.tii.ae/insights/tii-launches-falcon-emirati-alongside-new-arabic-ai-models-speech-and-visual-text
https://x.com/TIIuae/status/2107447340026245425
Hume AIModel providermodel release2026-10-06
Hume relaunches Expression Measurement API with new voice and face models
Builders who lost Expression Measurement after the June 2026 sunset get a limited-release path back just as Hume kills TTS and EVI on 13 November (2026-10-02-hume-tts-evi-api-eol), but the Gemini 3.8 Flash win is a vendor claim across 32 emotion datasets rather than a voice-agent latency or ASR bench.
Also See
Details3 sources · 2 X posts
Hume reopened its Expression Measurement API in limited release with new speech and face models that turn tone, delivery, and facial expression into searchable scores, claiming the speech model leads Gemini 3.8 Flash across 32 public emotion datasets.
Expression Measurement API
Someone says "I'm fine." You know they aren't.
You heard it - the pause, the flat tone. Voice AI doesn't. It all gets dropped at transcription.
So we built the Expression Measurement API. 48+ emotions, 600+ descriptors, 50+ languages, audio and video, real-time.
Conversation was always more than words.
https://x.com/hume_ai/status/2107496133388018034
Today we're relaunching Hume's Expression Measurement API with new models for voice and face, and our speech model beat Gemini 3.8 Flash on 32 public emotion datasets.
Transcripts capture WHAT people say. Expression Measurement captures HOW they say it. It turns tone, the delivery, the look on their face, turned into scores you can search, compare, and act on.
https://x.com/AlecInRealTime/status/2107507979226366312
FermionModel providermodel release2026-10-06
Fermion open-sources Phonon-2 for the Apple Neural Engine at 606× realtime
A first-party Neural Engine build lifts Phonon-2 from 174× to 606× realtime one week after its launch (2026-09-29-fermion-phonon-2) and leaves the GPU free for a local LLM beside it; the iOS 18 target opens the same weights to iPhone apps, though all speed figures are on Fermion's own M5 hardware.
Details3 sources · 2 X posts
Fermion published a Core ML package of its Phonon-2 English speech recogniser on Hugging Face with a Swift package and Python runner, and made it the default engine in its Detta Mac dictation app.
Phonon-2 Core ML
Detta now runs entirely on the Apple Neural Engine. An hour of speech in 6 seconds on a base MacBook Air, using 1/3rd the amount of memory as before with the same accuracy.
We're also open-sourcing the Core ML package for Phonon-2.
We're grateful for the response to Phonon-2, which has been the #1 trending speech recognition model on Hugging Face for the past four days. There is much more to come.
We're hiring across engineering and research, if you want to push what on-device Al can do, email careers@fermionresearch.com with your best work.
https://x.com/fermion_ai/status/2107510847170043939
We’re releasing the Core ML package for Phonon-2, reaching a staggering 606× real time, up from our 174× launch figure.
That’s an hour of speech in just 6(!) seconds, on a base M5 MacBook Air, with no loss in quality.
It’s now the default in Detta. We’ve cut peak app memory from 3.1 GB to 1.1 GB, and idle memory from 1.8 GB to just 600 MB, even with Gluon loaded!
That puts Detta's peak memory footprint same as WisprFlow, with the models in Detta running completely locally!
https://x.com/yoitsmanan/status/2107505508655759729
DeepgramModel providermodel release2026-10-06
Deepgram adds multilingual code-switching to Nova-3 Medical across 10 languages
Clinic and telehealth agents with mixed-language patients, such as Hindi and English callers on Deepgram's India endpoint (2026-09-15-deepgram-india-endpoint), can keep one medical model per call instead of routing by language; Deepgram gives no accuracy comparison with monolingual Nova-3 Medical.
Details2 sources
Deepgram's nova-3-medical model now accepts language=multi in batch and streaming, transcribing medical audio where speakers switch between Dutch, English, French, German, Hindi, Italian, Japanese, Portuguese, Russian, and Spanish.
Nova-3 Medical (multilingual)
Speaktext-to-speech10
LAIONModel providermodel release2026-10-11net-new
LAION open-sources the Humaneness Voice Converter and Acting TTS checkpoints under CC BY 4.0
LAION's own numbers show the converter trading speaker similarity (0.773 against 0.821 for untouched Chatterbox) for expressiveness, with no human listening or WER test, which makes it a research base for dataset augmentation rather than a cloning product; it adds a CC BY 4.0 option beside KittenTTS 2 (2026-10-02-kittenml-kitten-tts-2, W40) in open TTS.
Details5 sources
LAION posted a complete Chatterbox-based voice-conversion checkpoint and Small and Large voice-acting text-to-speech fine-tunes to Hugging Face as research releases for English and German speech.
Humaneness Voice Converter and Humaneness Voice Acting checkpoints
AnamPlatformlaunch2026-10-11
Anam adds ElevenLabs v4 voices, pronunciation dictionaries, and 44-language cloning
Interactive-avatar builders who already pick Anam for the face layer get the same ElevenLabs v4 expressiveness and brand-name IPA control that standalone TTS stacks shipped in late September (2026-09-28-elevenlabs-v4), without leaving the persona lab.
Also Show
Details3 sources
Anam Lab and the API can import ElevenLabs v4 or v4 Turbo voices, attach organisation pronunciation dictionaries with sounds-like or IPA entries, and clone Cartesia voices in 44 languages including Japanese, Korean, Hindi, and Urdu.
ElevenLabs v4 voices and pronunciation dictionaries · vendor ElevenLabs
HeyGenModel providermodel release2026-10-09
HeyGen Voice debuts as an in-house TTS model at #1 on Artificial Analysis Controlled Voice
Avatar builders who already use HeyGen no longer need a third-party TTS for the same identity stack: Instant Clone lands a HeyGen Voice id in seconds, and the 9 October Controlled Voice #1 ranking is Artificial Analysis preference data rather than a latency claim.
Also Show
Details5 sources · 2 X posts
HeyGen opened its in-house HeyGen Voice TTS on the product and API on 9 October, with Instant Clone from one recording and a temporary API rate of $15 per million characters through 31 October.
HeyGen Voice
Speechify streams completed sentences from LLM token input
The five-minute text-input cap and one-request concurrency constrain long-running or multiplexed sessions.
Also Connect
Details1 source
The `GET /v1/audio/stream/ws` endpoint converts LLM token streams into speech, emitting each sentence once complete.
WebSocket streaming endpoint
OcularPlatformlaunch2026-10-08
Ocular licenses studio-recorded voice-actor voices as default voices for TTS models
TTS vendors need cleared voices for their default line-ups, and a licensed catalogue with cloning rights moves Ocular from evaluation data into the TTS supply chain, next to its full-duplex datasets (2026-09-28-ocular-full-duplex-audiovisual-dataset).
Details2 sources · 1 X post
Ocular now sells a catalogue of professionally performed, studio-recorded voices that TTS builders can license as commercial default voices, with cloning rights included.
Default Voices for TTS
New Product: we're introducing Default Voices for Text-to-Speech (TTS) models.
A catalog of professionally performed, studio-recorded voices that can be licensed for commercial TTS default voice cloning.
Each voice comes with:
1. A professional voice actor
2. Studio-recorded
https://x.com/OcularHQ/status/2108319704607785076
Gradium Voice Localization moves an existing voice into another language or accent
Agents serving several markets can keep one brand or cloned voice across locales instead of casting a new voice per language; it builds on Voice Design (2026-09-07-gradium-voice-design) and arrived a day after the catalogue grew to 28 accents (2026-10-07-gradium-1000-new-voices).
Details2 sources · 1 X post
A new Gradium endpoint takes a flagship, cloned, or designed voice and returns candidates that speak a target language or accent, which developers audition and save as permanent voices.
Voice Localization
Developers can now localize existing voices to another language or regional accent while preserving its timbre, age, pace, and character.
Voice Localization is available through Gradium Studio and API.
https://x.com/GradiumAI/status/2110408068773384618
NineNineSix releases Akylai, a Kyrgyz and Kazakh TTS fine-tune of MOSS-TTS v1.5
Covers two Central Asian languages missing from the four in NineNineSix's own Gepard models (2026-07-07-nineninesix-gepard-1-0), and builds on a third-party base rather than Gepard; with 33 downloads by 17 October, uptake is still negligible.
Details2 sources
Open Apache 2.0 weights for voice-cloning speech generation in Kyrgyz and Kazakh, with Russian also evaluated.
Akylai
GradiumModel providerlaunch2026-10-07
Gradium adds more than 1,000 catalogue voices across 29 accents in five languages
Gradium built the batch with the Voice Design model it filed in September (2026-09-07-gradium-voice-design, W37), making this the first bulk output of that tool; the preference gains come from Gradium-run Elo comparisons rather than an external benchmark.
Details2 sources · 1 X post
Gradium's TTS voice catalogue grows from 387 to 1,422 voices for English, French, Spanish, German, and Portuguese, cast for customer service, narration, ads, and character use.
Gradium TTS voice catalogue
SierraPlatformlaunch2026-10-06
Sierra Persona Studio lets brands build an agent voice from their own recordings
August's Voice Personas (2026-08-07-sierra-voice-personas, W32) chose voices from Sierra's model constellation; uploading brand recordings takes Sierra into custom voice cloning, and the recap describes no consent or speaker-verification step. The 50% resolution gain comes from a single unnamed customer.
Also Orchestrate
Details1 source · 1 X post
Sierra customers can upload voice samples and tune an agent's mannerisms, formality, and style inside Agent Studio without engineering help.
Persona Studio
@btaylor @claybavor Meet Persona Studio, an all-new way to give your brand a unique voice. In our demo, our fictional outdoor retailer, Northwind Outfitters, creates their Scout agent with none other than @claybavor. https://t.co/dAVQdeAaIe
https://x.com/SierraPlatform/status/2107523483685286215
LatitudeEnterprisepartnership2026-10-05net-new
Latitude voices every Voyage narrator and character live with Inworld TTS
The two speed claims use different baselines—Voyage says four times faster, while Inworld says 3–5x—so they should not be combined into one benchmark; the six-week call counts indicate production scale but not quality.
Details1 source · 2 X posts
Voyage gives its narrator and every character a live generated voice, using World Engine delivery directions with Inworld TTS.
Voyage live narration on Inworld TTS · vendor Inworld
In Voyage, the narrator and every character you meet have their own voice, generated live for whatever you do.
Our SOTA World Engine gives every line of dialogue a delivery direction (warmly, breathless, urgently) and @inworld TTS performs it live.
Here’s what that unlocked:
4x faster voice at a quarter of the cost. Those savings helped us offer unlimited play to all our subscribers.
Our founder and CEO @nickwalton00 sat down with the Inworld team to talk about why a robotic voice breaks a story, and how we made every voice in Voyage feel real.
https://x.com/playVoyage/status/2107241058879029474
Key results:
- Voice at roughly a quarter of the cost of their previous provider
- TTS 3–5x faster than another provider in Latitude’s production tests
- More than 15 million TTS calls in six weeks
- More than 25 million inference calls over the same period
https://x.com/inworld/status/2112594816701735007
Orchestrateframeworks, evals14
Pipecat demo tracks character affect with a 28M-parameter Tone classifier
Multi-character voice agents get a CPU-side affect and conversational-move step (~10 ms) that can drive TTS delivery and character state without spending LLM tokens on every line — a companion to Audience's addressee routing (2026-10-09-pipecat-audience-classifier-demo) in the same SpellSpeak / Pipecat demo line.
Also Speak
Details2 sources · 1 X post
Jon Taylor shows SpellSpeak Tone, a ~28M-parameter ONNX MiniLM classifier that tags each transcript line for emotion, directed social acts, hostility, and conversational move inside a Pipecat multi-character agent, running in about 10 ms on CPU.
SpellSpeak Tone classifier demo
Here's Tone, another tiny decision model, fine-tuned from MiniLM and running in a Pipecat pipeline. It reads every line of a transcript for emotion, intent and hostility, and the harness uses that to track how a character feels about you as you talk. ~10ms on CPU.
In this demo, I'm using the output to pass tone guidance to a TTS service.
Confidence meters matter a lot for this model as it needs more training on sarcasm and subtle digs. For example, this character often responds with anger when I greet her with "Hello Nova, you're looking very orange today."
Model here: https://huggingface.co/spellspeak/tone
https://x.com/JonPTaylor/status/2108979064342028468
SierraPlatformpartnership2026-10-09
Sierra publishes Personal Agent Protocol draft Poppy with 35 new design partners
OpenAI and ElevenLabs joining the design table, plus Bank of America, Visa, and Wells Fargo, moves PAP from a Sierra Summit pitch with six launch partners to a draft that large consumer brands are willing to shape, while Decagon's PACT (2026-10-06-decagon-pact-open-source) remains a separate Apache-2.0 stack.
Also Connect
Details3 sources · 1 X post
Three days after the Meta partnership announcement, Sierra released a named draft of Personal Agent Protocol (Poppy) and added 35 design partners spanning banks, insurers, payments, and OpenAI and ElevenLabs.
Personal Agent Protocol (Poppy) draft
As @btaylor just announced we're publishing the first draft of Personal Agent Protocol, known as Poppy — and announcing 35 new design partners.
With Poppy, customers decide what access to give their personal agents, and companies set parameters for what those agents can do. It's built on open standards businesses already use, and designed to work through websites, APIs, and company agents.
This is a starting point — we welcome feedback from companies, developers, and personal agent builders to help shape this new protocol. Read more here: https://sierra.ai/blog/poppy or read the draft here: https://personalagentprotocol.org/.
https://x.com/SierraPlatform/status/2108665347201212623
Pipecat demo routes multi-character lines with a 33M-parameter Audience classifier
Multi-character voice agents get a CPU-side addressee step (~6–7 ms) that can answer unclear without every character's LLM reading the line — a sharper orchestrate-layer complement to the Jev routing in 2026-10-06-pipecat-unreal-multi-npc-demo and 2026-09-30-pipecat-multi-character-jev-handover-demo.
Details3 sources · 2 X posts
Jon Taylor shows SpellSpeak Audience, a ~33M-parameter ONNX classifier that decides who each spoken line is for inside a Pipecat multi-character agent, running in about 6–7 ms on CPU.
SpellSpeak Audience classifier demo
I’m obsessed with using tiny classifier models in Pipecat agents. This one, Audience, is a ~33M-param ONNX model built on Ettin that works out who each line in a scene is for. It uses names, descriptions, nearby items and where people are standing, and it can pick one character, several, the whole group, or return "unclear" if you're like me and talk jibberish.
It runs in ~6–7 ms on CPU and slots nicely into a Pipecat pipeline. Multi-character agents are really fun!
Weights here: https://huggingface.co/spellspeak/audience/tree/main
https://x.com/JonPTaylor/status/2108351556680634582
I’m obsessed with using tiny classifier models in Pipecat agents. This one, Audience, is a ~33M-param ONNX model built on Ettin that works out who each line in a scene is for.
https://x.com/kwindla/status/2108423640731836589
SanasPlatformlaunch2026-10-08
Sanas launches Supervised AI so one human can steer up to four voice agents mid-call
Sanas offers a third contact-centre pattern between agent-assist and fully autonomous bots: one supervisor for up to four AI calls with silent takeovers, invitation-only until a December 2026 wider release, and built on the same speech stack already pitched to healthcare and banking BPOs.
Also UseSpeak
Details3 sources
Sanas Supervised AI lets AI agents take customer calls while a supervisor watches up to four conversations, whispers private guidance, takes over mid-sentence, and hands control back without an audible transfer.
Sanas Supervised AI
Retell launches Conversation Memory so voice agents keep a running brief per contact
Native per-contact memory lands in Retell a week after Telnyx's Agent Memory API beta (2026-10-01-telnyx-agent-memory-api, W40), so builders on either stack get caller continuity without hand-rolled extract-and-store workflows; Retell's skippable Conversation Flow nodes are the productised twist, and the $0.005 backfill fee is the only published price.
Also Use
Details2 sources · 1 X post
Retell turns per-contact memory into a native toggle: after each phone call or SMS chat it rewrites a concise brief of preferences, past topics, and open follow-ups that later agents can use and conversation-flow nodes can skip against.
Conversation Memory
We're excited to launch Conversation Memory.
Until now, giving an AI voice agent memory meant building a workflow to extract information after each call, store it, and retrieve it for the next conversation. The more context you wanted your agents to remember, the more fields, mappings, and logic you had to configure and maintain.
Today, we're bringing that work natively into Retell.
Turn memory on with a simple toggle, and Retell maintains an evolving brief for each contact, including relevant preferences, what's been discussed, and what still needs follow-up. As conversations progress, the brief updates, carrying useful context forward and clearing resolved action items.
That context can also shape the path through your Conversation Flow. With skippable nodes enabled, agents can skip questions when the required information is already available and focus on what still needs an answer.
A customer following up on a replacement has already explained the problem. Their next call should pick up where they left off and move the request forward.
Learn how to get started:
https://www.retellai.com/blog/ai-voice-agent-memory
https://x.com/retellai/status/2108248018080182761
LiveKit Agent Simulations reaches GA for Cloud projects with text and audio scenario runs
LiveKit now ships first-party whole-conversation eval beside its Agents runtime, so Cloud customers can regression-test voice paths without bolting on a third-party harness; free-through-October pricing and the Serve AI 98% case study are vendor-framed, and Evaluations is still roadmap.
Details5 sources · 1 X post
LiveKit opens Agent Simulations to every LiveKit Cloud project: LLM-driven simulated callers run full scenarios against the real agent in text or audio, with an LLM judge scoring pass or fail.
Agent Simulations
Know exactly what your agent can do, catch gaps before customers do, and test any model against your own scenarios before you switch. Your agent has a QA team now.
Try LiveKit Simulations free through October: https://livekit.com/products/agent-simulations
https://x.com/livekit/status/2108265592348422525
CiscoPlatformlaunch2026-10-07
Cisco adds the Dialog harness to Webex AI Agent and lines up Nemotron ASR and Nova 2.5 Sonic for contact-centre voice
Amazon's Nova 2.5 Sonic (2026-10-05-amazon-nova-2-5-sonic) gains a second downstream platform after LiveKit Agents 1.8.5 (2026-10-06-livekit-agents-1-8-5-think), but Webex customers wait for a December 2026 early-access window, and Dialog itself is a Q1 2027 beta, so most of this release is roadmap.
Also ThinkHearConnect
Details3 sources
At WebexOne in Austin, Cisco introduced Dialog, a harness that carries customer context across channels and keeps Webex AI Agent working on a case after the conversation ends, with new speech-model and third-party voice-agent options for Webex Contact Center.
Webex AI Agent (Dialog) · vendor NVIDIA Nemotron and Amazon Nova 2.5 Sonic
Automation Anywhere agrees to buy Boost.ai for customer-facing conversational and voice AI
An automation vendor buying a European voice-and-chat agent platform is the third voice M&A deal in eight days, after Inworld and Ultravox (2026-09-30-inworld-acquires-ultravox) and TCN and Abstrakt (2026-10-01-tcn-acquires-abstrakt); the deal has not closed, and the 90%+ resolution rate is Boost.ai's own figure.
Details1 source
Automation Anywhere signed a definitive agreement to acquire Boost.ai from Nordic Capital, adding Boost.ai's voice and chat agents for banks, telcos, and insurers to its automation platform so a customer request can run through to back-office execution.
Boost.ai conversational and voice AI platform
SierraPlatformpartnership2026-10-06
Sierra and Meta propose Personal Agent Protocol for consumer agents dealing with businesses
Decagon published its own personal-agent discovery and authorisation spec with Personal Agent Gateway five days earlier (2026-10-01-decagon-personal-agent-gateway), so two CX platforms now back separate specs for the same Muse traffic. Phone calls are one route the protocol could replace, but the post names no telephony or voice-specific handling and no specification text exists yet.
Also Connect
Details4 sources · 2 X posts
Sierra and Meta, with Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart, are drafting an open standard for how consumer agents such as Muse sign in to a company and reach its website, APIs, or company agent.
Personal Agent Protocol
Today we're announcing Personal Agent Protocol — an open standard @Meta and @SierraPlatform are developing along with industry partners at @Genesys, @instinct, @RocketOTD, @Shopify, @stripe, and @Walmart. It will help define how personal agents interact with businesses and is https://t.co/Pr2JXFka3P
https://x.com/SierraPlatform/status/2107531483837862324
Today we're announcing Personal Agent Protocol — an open standard @Meta and @SierraPlatform are developing along with industry partners at @Genesys, @instinct, @RocketOTD, @Shopify, @stripe, and @Walmart. It will help define how personal agents interact with businesses and is open for anyone to implement. You can read more here - and if anyone is interested in joining let me know! https://sierra.ai/blog/introducing-personal-agent-protocol
https://x.com/btaylor/status/2107535182119244176
SierraPlatformmodel release2026-10-06
Sierra adds Curie orchestration and Fleming agent-caller detection models to Agent OS
Fleming is the third agent-caller detector in nine days, after Liberate's AI Intercept (2026-09-28-liberate-ai-intercept) and Decagon's Personal Agent Gateway (2026-10-01-decagon-personal-agent-gateway), all aimed at Muse-class assistants that place calls (2026-09-17-meta-muse-outbound-calls, W38). None of the three has published a detection rate.
Details2 sources · 2 X posts
Two new Sierra-tuned models join Agent OS: Curie runs the conversation loop of path choice, tool calls, and replies, and Fleming flags when a phone caller is another AI agent.
Agent OS: Curie and Fleming
ElevenLabs adds Architect, a built-in assistant that builds and tunes ElevenAgents by voice or chat
Sierra showed its reimagined Ghostwriter teammate (2026-09-28-sierra-ghostwriter-teammate, W40) at its Summit the same day, and Decagon's Duet Apprentice (2026-10-01-decagon-duet-apprentice) drafts AOPs from Slack and docs, so three CX agent vendors now ship an AI that edits the voice agent itself. ElevenLabs publishes no resolution-rate figures behind its 'week over week' framing.
Details2 sources · 1 X post
ElevenAgents customers can ask Architect to build a new agent from a description and attached files, or to diagnose transcripts and return simulation-tested fixes for review.
ElevenAgents Architect
Decagon open-sources PACT and joins the Personal Agent Protocol working group
Decagon's seat in the Personal Agent Protocol group, taken on the day Sierra and Meta announced it (2026-10-06-sierra-meta-personal-agent-protocol), puts both CX vendors behind one standards effort for Muse-class traffic; how PACT's A2A scopes fit PAP's promised v0.1 spec is not stated.
Also Connect
Details3 sources · 1 X post
Decagon released PACT, an Instinct co-developed protocol that lets a brand's support agent check which personal agent is calling and what the customer authorised it to do, under Apache 2.0, and joined the Muse-led Personal Agent Protocol working group.
PACT (Personal Agent Consent & Trust Protocol)
Today, we're open-sourcing the Personal Agent Consent & Trust Protocol (PACT), co-developed by @Instinct. We're also joining the Personal Agent Protocol working group by @Muse.
Built on A2A and OAuth, PACT enables businesses to verify who a personal agent is acting for and what the customer has authorized it to do.
Excited to shape the industry standard for how personal agents will interact with businesses.
https://x.com/DecagonAI/status/2107544569340694705
telliPlatformlaunch2026-10-05
telli adds Browser Call testing to its voice-agent builder
Testing in the browser removes the phone-number step from agent iteration and pairs with the Week 42 simulations (2026-10-12-telli-simulations); API users must move off call_outcome and call_analysis before 1 February 2027.
Also Connect
Details2 sources · 1 X post
The test-call dialog lets builders use a microphone and follow the transcript live without placing a phone call.
Browser Call
🚀 Day 2/5 of telli Launch Week
Introducing Web Calls: test your voice agent directly in telli, without picking up the phone.
Talk to your agent and follow the transcript as it happens.
And when real customers call, you can listen in and follow those conversations live, too. https://t.co/OSaAqab2xe
https://x.com/tellicom/status/2112444335224754527
OcularPlatformfunding2026-10-05
Ocular raises a $2M pre-seed led by Drive Capital for voice training data and benchmarks
Full-duplex speech models need separate-channel conversation with real overlaps, which barely exists on the open web, and Ocular's claimed seven-figure revenue at pre-seed suggests labs already pay for it.
Also HearThink
Details2 sources
Ocular AI, which records conversational speech and builds evaluations for labs training voice-native models, raised $2 million led by Drive Capital with Y Combinator, Alumni Ventures, 1745 Ventures, Orange Collective, MyAsia VC, and angels.
Ocular AI
Useagents in market40
ToriPlatformlaunch2026-10-11net-new
Tori turns your own Japanese sentences into listening drills and AI speaking practice
ChickyTutor (2026-10-08-chickytutor-whiteboard-hybrid-french) is a spoken tutor across many languages with a whiteboard; Tori is Japanese-only and mines the learner's own sentences into TTS listening loops before the speaking tutor — same education layer, different input path.
Also HearSpeak
Details2 sources
Investiture Labs' Android app builds personal sentence islands with spoken audio, then drills them through screen-off listening, shadowing, AI stories, SRS flashcards, and an AI tutor across 15 scenarios.
Tori
HireKeyPlatformlaunch2026-10-11net-new
HireKey ships a live AI voice coach for STAR interview practice
Candidate-side spoken practice lands days after ParakeetAI 2.0's mock interviewer (2026-10-07-parakeetai-2-0); HireKey pairs the voice coach with role-fit and resume tools rather than a live employer-call copilot, and the Microsoft/Amazon/Stripe recruiting claim is the makers' own.
Also Hear
Details2 sources
HireKey's job-search co-pilot adds an on-demand spoken interview coach that asks follow-ups and gives STAR-method feedback, beside resume fit scoring, tailoring, and application tracking.
HireKey
FoleyPlatformlaunch2026-10-11net-new
Foley launches a voice to-do list that files tasks and asks before it guesses
Spoken capture into a full task system with confirmation-before-guess sits beside dictation apps such as Featherweight (2026-10-07-featherweight-dictation-launch) that only type at the cursor; Foley plans the day from the same spoken list on watch and phone.
Also Hear
Details3 sources · 2 X posts
Foley turns spoken captures into dated, project-tagged tasks across phone, watch, desktop and web, and asks when a project or time is unclear instead of inventing one.
Foley
Voiceflex launches VoiceAI, a white-label phone agent for UK channel partners
A channel-only UK carrier now resells a white-label voice agent onto the trunks partners already sell — no stack, pricing, or live customer is named on 9 October, so treat it as a distribution ship rather than a model launch.
Also Connect
Details3 sources
UK SIP and hosted carrier Voiceflex opened VoiceAI, a partner-only white-label voice agent that answers, verifies, books, and routes calls on the customer's existing telephony, including Microsoft Teams.
VoiceAI
Phonable replaces iPhone carrier voicemail with miss-call texts, transcripts, and AI summaries
Consumer missed-call handling sits beside personal outbound callers such as ClawCall (2026-10-05-clawcall-chatgpt-plugin) and Tab (2026-10-06-tab-personal-agent-calling), but Phonable only covers the inbound voicemail leg on iPhone and never names its STT or TTS stack.
Also HearSpeak
Details3 sources
Phonable forwards unanswered iPhone calls to its own line, texts the caller why you missed them, and turns any voicemail into a transcript plus a short AI summary, with switchable modes and recorded or AI-spoken greetings.
Phonable
EliseAIPlatformpartnership2026-10-09
EliseAI joins Nextech as a preferred partner for specialty-practice Voice AI
The joint offering is dated for customer availability at the start of 2027, so the 9 October wire is a preferred-partner and integration announcement rather than a go-live, and Nextech practices still choose between EliseAI and Assort Health on the same EHR stack.
Also Connect
Details2 sources
EliseAI becomes a Nextech preferred partner for inbound and outbound patient calling in ophthalmology and dermatology, with joint APIs, implementation, and support aimed at a start-of-2027 customer launch.
Nextech preferred partnership (Voice AI patient access) · vendor Nextech
Assort Health becomes a Nextech preferred partner for specialty-practice voice access
A second specialty EHR preferred seat next to NextGen widens Assort's distribution into ophthalmology, dermatology, plastic surgery, and med spa books that already run Nextech, while Nextech frames the deal as choice against EliseAI rather than a sole-source lock-in.
Also Connect
Details2 sources
Nextech named Assort Health a preferred agentic voice partner so specialty practices can run Assort's inbound and outbound patient-access agents inside Nextech scheduling, intake, and readiness workflows.
Nextech preferred partnership (agentic patient access) · vendor Nextech
The University of Texas System opens Abridge to all eight UT health institutions
Keeping deployment decisions local makes the deal a shared access framework rather than proof that all eight institutions have gone live; it lands two weeks after Abridge's VA enterprise contract path (2026-09-22-abridge-va-enterprise, W39).
Also Hear
Details2 sources · 1 X post
An enterprise-wide agreement gives all eight UT health institutions systemwide access to Abridge through the UT REAL Health AI programme, with each institution retaining deployment decisions.
Abridge enterprise-wide agreement · vendor Abridge
It’s a privilege to partner with the @utsystem on an enterprise-wide agreement. Eight UT health institutions gain systemwide access to Abridge, with decision-making kept local.
Dr. Peter McCaffrey in @FierceHealth on what it means for clinicians and patients:
https://x.com/AbridgeHQ/status/2112543023124901953
Trasti signs ElevenLabs to replace insurance purchase forms with a voice assistant
Admiral uses an ElevenLabs-built agent for loan settlement quotes (2026-08-10-admiral-elevenagents-settlement-quotes); Trasti puts the same vendor at the point of sale for a regulated product. Like the Karnataka MoU (2026-10-06-karnataka-elevenlabs-voice-ai-mou) the same week, it has no launch date.
Details1 source
HPI GMA, which runs the Polish online insurer Trasti, signed a strategic deal for ElevenLabs voice agents that will talk customers through choosing and buying a policy instead of filling in a form.
Voice insurance sales assistant · vendor ElevenLabs
Pegasus Airlines invests in Zingly.ai and puts agentic voice AI on guest-service calls
A major European low-cost carrier is putting money and production traffic behind an agentic voice vendor rather than a chatbot-only path: the first use case is live in Turkish with nine more journeys planned by mid-2027, though deal size and containment rates are unpublished.
Also Think
Details3 sources
Pegasus Innovation Lab becomes an investor and design partner in Zingly.ai, with the first inbound guest-service voice use case live and a roadmap through mid-2027 covering the top ten contact-centre journeys.
Agentic voice guest service · vendor Zingly.ai
Dialog Axiata launches Travel Sri Lanka, a multilingual tourism voice line for Southern Province travellers
A national mobile operator is putting a tourism FAQ on a plain PSTN number with 12+ languages and no app requirement—the Southern Province pilot on 011 778 8288 is the first Dialog AI line filed in this corpus, though the speech stack and any containment figures stay unpublished.
Also Connect
Details2 sources
Dialog and the Ruhunu Tourist Bureau open a phone-only AI travel assistant on 011 778 8288 that answers in more than 12 languages with no app or data connection, starting as a Southern Province pilot.
Travel Sri Lanka
ChickyTutor adds an interactive whiteboard to its spoken language tutor and pairs with an Amsterdam French school
Aristotle's voice tutor (2026-09-16-aristotle-voice-tutor, W38) also pairs speech with a whiteboard, for school subjects; ChickyTutor's Amsterdam course from 7 December places the AI between human-taught classes rather than in place of them.
Details3 sources
ChickyTutor, a voice conversation tutor for more than 70 languages, now draws whiteboard explanations and exercises during spoken practice, has an Android app, and supplies daily AI speaking practice for L'école de français hybrid courses.
ChickyTutor
Backbase opens Conversational Banking GA so bank AI agents can finish voice and text requests under bank-set controls
A banking OS vendor is selling governed voice-and-text agents as a GA product rather than a pilot slide: more than ten production banks and a vendor-claimed jump from 20% to 70% containment at one large South African bank are the published proof points, with no speech stack named.
Also Orchestrate
Details4 sources
Backbase's Banking OS agents take natural-language voice and text requests such as card block-and-replace or payment checks, run only the actions the bank permits, and hand incomplete work to employees with the context already gathered.
Conversational Banking
Underheard opens an AI voice interviewer that calls customers and ranks what to fix
Outbound interview agents sit beside Tab (2026-10-06-tab-personal-agent-calling) and ClawCall (2026-10-05-clawcall-chatgpt-plugin) as systems that place calls for a user, but Underheard is pitched at research and post-stay feedback with ranked themes and CRM write-back rather than booking errands, and it names no speech stack.
Also ThinkConnect
Details5 sources · 1 X post
Zurich-based Underheard reached general availability with a voice agent that designs qualitative customer interviews, places the calls or chats, and returns ranked themes with quotes and CRM write-back within minutes.
Underheard
We're live on Product Hunt today! 🚀
We built Underheard to help teams hear from more customers through AI-led interviews.
Take a look, tell us what you think, and if you like what we're building, we'd appreciate your upvote.
https://t.co/vgoVQXQcjp
https://x.com/underheard_ai/status/2108478561111699566
TwinsAIPlatformlaunch2026-10-07net-new
TwinsAI ScreenPass answers phone call screeners on behalf of outbound sales reps
Outbound callers are now fielding agents to talk to consumers' screening agents, the reverse of Liberate AI Intercept (2026-09-28-liberate-ai-intercept, W40) and Sierra Fleming (2026-10-06-sierra-curie-fleming-models), which detect calling agents on the business side; the 28% screener rate comes from TwinsAI's own call logs.
Details2 sources
TwinsAI added ScreenPass to its AI parallel dialer: when an outbound sales call is picked up by an AI call screener on iPhone or Android, ScreenPass answers the screener for the rep and handles the call until a person comes on the line.
ScreenPass
ParakeetAI 2.0 adds a spoken AI mock interviewer to its live interview copilot
A candidate-side counterpart to recruiter screening agents such as Fireflies Voice Agents (2026-09-02-fireflies-voice-agents, W36); the live Call Assistant drafts answers during real employer interviews, and the 1.5M-user figure is ParakeetAI's own.
Also Hear
Details2 sources
ParakeetAI added a company-by-company Question Bank and voice mock interviews that ask, listen, and follow up to its Call Assistant, which listens to a real job interview and drafts answers live.
ParakeetAI 2.0
Instinct partners with Sierra and Decagon so its personal agent can settle support issues without a phone call
Instinct Concierge (2026-09-16-instinct-concierge-calling, W38) placed phone calls for users three weeks ago; this deal gives the Personal Agent Protocol (2026-10-06-sierra-meta-personal-agent-protocol) and PACT (2026-10-06-decagon-pact-open-source) their first named consumer-agent client, which could take some support traffic off the voice channel once it ships.
Also Orchestrate
Details3 sources · 2 X posts
When an Instinct user's personal agent reaches a business that runs Sierra or Decagon, the two agents will work the problem together, for jobs such as Wi-Fi troubleshooting, delayed packages, higher bills, and return or cancellation policies.
Instinct native customer service
Infinitus launches FieldForce agents that take provider calls for pharma sales and reimbursement teams
Infinitus takes its phone agents from payer and hub calls (2026-07-30-infinitus-ai-first-hub, W31) into pharma commercial teams, where the hand-off to Medical Information is the compliance line buyers will test; no pharma customer is named at launch.
Also Orchestrate
Details2 sources
FieldForce gives pharmaceutical sales representatives and field reimbursement managers AI agents that prepare visits, update the CRM, and answer providers' inbound calls around the clock between visits.
FieldForce
Highmark Wholecare adds Fortuna Health's phone agent to Medicaid renewal support in Pennsylvania
A Medicaid plan is putting a voice agent on renewals before the 1 January 2027 federal changes, in a programme where a missed call can cost a member coverage; the 9-point renewal lift is Fortuna's own comparison, and the agent is the one Fortuna launched in February on OpenAI, ElevenLabs, and Deepgram (2026-02-19-fortuna-health-voice-ai, W08).
Details2 sources
Wholecare members due to renew Medicaid will get bilingual calls from Fortuna's voice agent that explain deadlines and next steps, answer common questions, and pass complex cases to live navigators.
Fortuna Voice AI for Medicaid renewals · vendor Fortuna Health
General Dynamics Land Systems teams with Primordial Labs to put Anura voice control on Abrams and XM30 vehicles
Voice command moves from Primordial's small-drone demonstrations onto heavy armour just as XM30 plans to cut the crew from three soldiers to two; this is a teaming agreement with no fielding date, and Anura is kept away from lethal tasks by design, so it reads as a crew-workload tool.
Details3 sources
A master teaming agreement brings Primordial Labs' Anura, which turns crew voice commands into actions on sensors, displays, radios, and uncrewed systems, to vehicles across the GDLS portfolio.
Anura natural language command and control · vendor Primordial Labs
FlylePlatformfunding2026-10-07net-new
Flyle raises a ¥1.88B Series A and opens AI Operator for Japanese contact centres
Japan's contact-centre AI market gets another funded native stack: Flyle's ¥1.88B Series A pairs the raise with an AI Operator soft launch ahead of a November GA, though the release gives no resolution rates or speech-vendor stack.
Also Orchestrate
Details4 sources
Tokyo-based Flyle closed an ¥1.88 billion Series A that lifts cumulative funding to about ¥2.27 billion, and began offering AI Operator agents that take phone, chat, and email requests through to system actions.
Flyle AI Operator
eDreams ODIGEO brings flight, hotel, and car-rental booking to Alexa+ over MCP
Alexa+ reaches a travel seller's inventory through that seller's own MCP network rather than a bespoke skill, the first third-party Alexa+ booking integration in the corpus; eDO names no launch date or markets, so treat it as an announced collaboration rather than a measured go-live.
Details2 sources
eDreams ODIGEO connected its travel subscription platform to Amazon's Alexa+ through its own Model Context Protocol network, so Alexa users can search for and book flights, hotels, and car rentals by voice.
Travel search and booking on Alexa+ · vendor Amazon (Alexa+)
Desible.ai raises a ₹32 crore seed plus round led by Prime Venture Partners for BFSI voice agents
A second Indian BFSI voice-agent round in five weeks after Navana.ai's ₹40 crore Series A (2026-09-07-navana-series-a, W37), with Desible selling compliance rules (IRDAI, RBI, and DPDPA) rather than voice quality; the 10 million monthly engagements are unaudited company figures.
Details3 sources
Desible.ai, a Bengaluru start-up whose voice agents handle renewals, collections, claims intimation, and welcome calls for Indian insurers and lenders, raised a seed plus round led by Prime Venture Partners with existing backer Invention Engine.
Desible.ai BFSI voice agents and workflow orchestration
CiscoPlatformlaunch2026-10-07
Cisco unveils Personal Agents for Webex Calling that screen and act on employee phone calls
Webex moves contact-centre agent tech onto every employee's desk phone a day after Dialog targeted CX journeys (2026-10-07-cisco-webex-dialog-voice); both are still roadmap into 2027, and Computerworld says Personal Agents will sit behind a paid Enhanced AI add-on rather than the base Suite.
Also Orchestrate
Details4 sources
At WebexOne 2026 Cisco introduced personal agents for individual Webex Calling users that answer when the employee is busy, judge urgency, and take actions such as scheduling a follow-up or capturing CRM notes.
Personal Agents for Webex Calling
AperoPlatformfunding2026-10-07net-new
Apero raises a $5M Series A from Now Book It for its bilingual restaurant phone receptionist
The investor is a Valsoft sister company, so the $5M is a reservation-software group backing phone booking inside its own book; Square's kiosk voice ordering (2026-10-14-square-kiosk-voice-ordering, W42) goes after the same restaurants from the counter.
Details4 sources
Now Book It, part of Valsoft, invests in the Montréal reservation platform whose 24/7 AI receptionist takes bookings by phone in French and English.
Apero 24/7 receptionist
VoccaPlatformfunding2026-10-06net-new
Vocca raises a $20M Series A led by Norrsken VC for its medical-practice phone assistant
Front-desk phone automation for clinics is drawing Series A money in Europe as well as the US: Vocca's round follows InTouchNow's £2.3M NHS GP seed (2026-09-21-intouchnow-seed-funding, W39), though Vocca's 70% resolution rate is self-reported and no customer is named.
Details4 sources
The Paris- and New York-based startup, whose assistant answers patient calls and books and reschedules appointments for medical practices, closed a $20 million Series A.
AI medical receptionist
TabPlatformlaunch2026-10-06net-new
Tab leaves stealth with a texting assistant that phones businesses and pays bills for its users
Tab joins Meta Muse (2026-09-17-meta-muse-outbound-calls, W38) and Instinct Concierge (2026-09-16-instinct-concierge-calling, W38) as a consumer agent that places its own calls, adding to the agent-caller traffic Sierra's fleming-1 (2026-10-06-sierra-curie-fleming-models) is built to flag; with no disclosed round behind the $300M valuation, it is filed as a launch rather than funding.
Also Connect
Details3 sources
Tab is a private-beta assistant that people text on iMessage or WhatsApp; it has its own computer, browser, phone number, and wallet, and uses them to book, buy, and call businesses on the user's behalf.
Tab
SesamePlatformlaunch2026-10-06
Sesame opens its voice assistants to all users with their own computers and app connections
Sesame frames these assistants as the voice OS for glasses due in 2027, so the app is the first public test of its voice-first agent delegation; it follows the 39-country iOS preview in W22 (2026-05-27-sesame-ios-preview) and the TurnBench release in W35 (2026-08-28-sesame-turnbench).
Also ThinkSpeak
Details3 sources · 2 X posts
Sesame's October release puts Maya, Miles, Charlie, and Simone in the iOS and Android Sesame app for everyone, each with its own computer for research and longer tasks, Gmail and Calendar connections, user-supplied MCP servers, and a new voice model.
Sesame voice assistants (October release)
Frames you love to wear, with an assistant you enjoy talking to.
Voice assistants available to everyone today. Glasses coming 2027.
This release is a step towards our vision—a voice OS on the frames that can connect to your tools, act on your behalf, and help you throughout your day.
iOS: https://ses-a.me/kdq6join
Android: https://ses-a.me/tessiois
https://x.com/sesame/status/2107535948041060713
With our latest models, Maya, Miles, Charlie, and Simone are smarter and more capable than ever:
- Each assistant has its own computer. They can riff with you on a personal project, research your next adventure, or help you find anything on the web.
- They can connect with apps like Gmail and Calendar to act on your behalf.
- A new voice mode for all characters making them more expressive.
- Artifacts that are now more helpful with new formats, including voice notes.
- Everything can be handled smoothly over voice or text.
Learn more: https://ses-a.me/8y5tku91
https://x.com/sesame/status/2107536741062373604
Rhem Labs opens preorders for Rhem, a talking tabletop robot that phones businesses for older adults
Rhem brings the consumer outbound calling of Google Gemini Call for Me (2026-09-24-google-gemini-call-for-me, W39) to an elder-care device, a different route from My Friend's phone-call companion (2026-09-23-my-friend-phone-companion, W39); with hardware due on 20–31 January 2027, the calling agent is a preorder promise for now.
Also Connect
Details2 sources · 1 X post
YC F26 startup Rhem Labs launched Rhem for ageing parents, who talk to the device to set reminders, ask questions, and have its agents call pharmacies or restaurants and complete online tasks, while family members follow readings and requests in an app.
Rhem
my dad and i built a robot for your parents.
today we’re launching @rhem_labs
our family spent nearly 20 years caring for my grandmother and aunt at home. we saw small health changes go unnoticed between check-ins. and simple things, like making a call or booking an appointment, meant waiting for someone else to help.
we built the robot we wish they’d had.
your parent just talks to Rhem. it helps make calls, book appointments, manage reminders, and get everyday things done online.
through the family app, you can follow changes in health readings and routines, see what Rhem handled, and step in when you’re needed. your parent chooses what to share.
Preorder Rhem Today → rhem.ai
https://x.com/jacobshenn/status/2107513493310337130
NEC Nexa Solutions launches a DEN.Ai phone agent for Japanese firms and municipalities
A domestic SI now wraps a startup phone agent in its own cloud and support for Japan's mid-market and municipalities, a resale pattern similar to BT Business with RingCentral (2026-10-01-bt-ai-receptionist, W40); NEC Nexa gives no pricing, latency, or customer count.
Also Connect
Details2 sources
NEC Nexa Solutions began offering a managed conversational phone agent, built on Decoding X's DEN.Ai and hosted on NEC Nexa's cloud, that takes first-line calls, captures the caller's request, and transfers to staff when needed.
Conversational Voice AI Agent Powered by DEN.Ai · vendor DEN.Ai (Decoding X)
Karnataka signs a voice AI MoU with ElevenLabs covering citizen services and voice restoration
The MoU is exploratory and names no live citizen service, unlike ANGRAU's Telugu farm-advice line in neighbouring Andhra Pradesh (2026-10-05-angrau-riaa-telugu-farm-voice-agent), which already takes calls. The India spending figure is a CEO interview statement, made while Sarvam sells Indian-language voice agents at ₹3.50 a minute (2026-08-13-sarvam-voice-agents-ga, W33).
Also Speak
Details3 sources · 1 X post
The state's Karnataka Innovation and Technology Society will work with ElevenLabs on voice access to government services in Kannada and other local languages, free voice-restoration licences for people who have lost speech, and safeguards against voice cloning and fraud.
ElevenLabs voice AI MoU (KITS) · vendor ElevenLabs
At the ElevenLabs Summit in Bengaluru, we signed our first MOU with an Indian state government to pilot voice AI.
Over 70% of the 100M+ ElevenAgents conversations in India over the past year were in Hindi, Kannada, Tamil, and Telugu.
Voice AI in India is multilingual, and the Summit brought together 500+ enterprise leaders, builders, and partners who are building for this future.
https://x.com/ElevenLabs/status/2107669995698246044
InvocaPlatformlaunch2026-10-06
Invoca launches home services AI agents that book jobs in ServiceTitan and Salesforce Field Service
Google began billing Local Services Ads advertisers for some unanswered calls on 1 October, and Invoca pitches its voice agent to plumbers and HVAC firms as the fix; it competes with ServiceTitan's own AI Virtual Agent and names no customer running it yet.
Also Connect
Details3 sources
Invoca released a home services package whose voice and messaging agents answer calls, texts, and web leads around the clock, check service area and intent, and book, reschedule, or cancel jobs in ServiceTitan or Salesforce Field Service.
Home services solution (AI voice and messaging agents)
Incredible relaunches its desktop agent as a hold-to-talk assistant that clicks and types in your own apps
The August funding post (2026-08-18-incredible-pre-seed-funding, W34) promised a public release later in 2026; this version acts inside the user's own logged-in sessions, whereas Sesame's assistants (2026-10-06-sesame-voice-assistants-agents) each run on a separate computer of their own.
Also See
Details2 sources
Swedish startup Incredible rebuilt its Mac and Windows assistant around a single key: hold fn, say the task, and it reads the screen and works through the user's signed-in browser, local files, and desktop apps.
Incredible
FlaiPlatformfunding2026-10-06net-new
Flai raises a $27M Series A led by Base10 for dealership call and scheduling agents
Auto retail joins this autumn's vertical voice funding run with dealer groups (Friedkin and Findlay) and Toyota's venture arm on the cap table; Flai pitches dealership depth against general small-business agents such as Meta's Muse (2026-09-17-meta-muse-outbound-calls), and the 50,000 monthly bookings are unaudited company figures.
Details2 sources
Flai, whose AI answers car-dealership sales and service calls, texts, and email and books appointments, raised $27M led by Base10 Partners with Toyota Ventures, First Round Capital, Y Combinator, The Friedkin Group, and Findlay Automotive Group, and introduced an AI-native Service and Sales CRM.
Flai (dealership AI agents and CRM)
CreditNirvana launches a voice co-browsing assistant and a 200-agent collections orchestrator for Indian lenders
Indian lending collections gets a second voice-agent suite within a week of Credgenics's Prix AI (2026-10-03-credgenics-prix-ai, W40); CreditNirvana pairs speech with on-screen co-browsing, and its 40% cost and 60–70% autonomy figures are vendor claims with no named lender.
Also Orchestrate
Details2 sources
The Perfios-owned debt-collection vendor released CognifAI Assist, a multilingual voice assistant that walks borrowers through lending screens with visual co-browsing, alongside ORCA, which picks the next collections action per account across digital, voice, field, and legal channels.
CognifAI Assist and ORCA
Slang AI relaunches its restaurant phone agent as Superhost with guest memory and 97 languages
Restaurant voice agents move from reservation answering to CRM-style guest memory: recalling allergies and visit history is now a vendor-stated feature across 97 languages, though the release gives no rollout schedule per capability beyond 'over the coming months'.
Also Think
Details2 sources
Slang AI is rolling out a rebuilt Superhost voice agent to full-service restaurant partners that remembers regulars, handles interruptions, books OpenTable Experiences, and lets operators tune accent and personality.
Superhost
Molin adds Phone AI so its Shopify shop assistant answers calls
Chat-first commerce assistants are adding the phone leg without a new plan: Molin switches it on for 1000+ shops it already serves, which puts voice beside the chat widget for small European and Shopify merchants rather than as a separate contact-centre purchase.
Also Connect
Details3 sources
Molin AI opened Phone AI to every e-commerce shop on its platform at no extra charge, so the agent trained on a store's catalogue, orders, and policies now answers calls on a Molin-provided or existing business number.
Phone AI
Florida Prepaid puts Agentforce Voice first on 75% of business-hours calls
A regulated state college-savings board reached 75% first-contact coverage with a knowledge-only agent that never touches account data, a useful ceiling for unauthenticated voice deployments; the 1.6- versus 10.9-minute handle times come from a Salesforce-sponsored piece with no go-live date.
Also Connect
Details1 source
Florida Prepaid's Agentforce Voice agent, reached through Amazon Connect, answers enrolment, eligibility, and plan-usage questions from approved knowledge articles and passes sensitive calls such as cancellations to staff with an AI summary.
Agentforce Voice (contact centre) · vendor Salesforce · delivered by NeuraFlash
ClawCall lists its phone-calling plugin in ChatGPT's directory, giving ChatGPT and OpenAI dots a US phone line
ClawCall sells calling as a $4.99 add-on to assistants people already use, whereas Meta Muse (2026-09-17-meta-muse-outbound-calls, W38) and Instinct Concierge (2026-09-16-instinct-concierge-calling, W38) build it into their own agents; Vapi's Personal Agent Calling (2026-09-30-vapi-personal-agent-calling, W40) chases the same assistants from the developer side, and ClawCall's coverage stops at English calls to US numbers.
Also Connect
Details3 sources · 3 X posts
ClawCall's official ChatGPT plugin lets users ask ChatGPT to call a US business, which ClawCall dials, steers through the phone menu and hold queue, and reports back on with any transcript it captured.
ClawCall for ChatGPT
ANGRAU launches RIAA, a Telugu farm-advice voice agent on ordinary phone calls
Public-sector farm advice in India is moving onto plain voice calls aimed at farmers without data connections; the 5,000-calls-a-day capacity is the only figure given, and the report names no model stack.
Details2 sources
Andhra Pradesh's agriculture minister launched RIAA for Acharya N. G. Ranga Agricultural University (ANGRAU), a 24/7 generative voice agent that advises farmers on paddy cultivation, fertiliser use, and pest management over regular calls in Telugu.
RIAA (Rythu Mitra Salahavani) · vendor Farm Vaidya