hotwirednews Get GrokBot

Week 43

19–25 Oct 20262026-W43 · 42 filings · 35 companies · 12 net-new

In progress. New filings arrive every 6 hours.

The Use layer leads so far with 11 of 42 filings, from Channel Corporation, Yelp, wevnal, SpeakON and 7 more.

Hearspeech-to-text6

RimeModel providermodel release2026-10-20

Rime launches Rime STT, its first speech-to-text model, for English and Spanish

Rime now sells both ends of a cascaded voice stack, from Mist v3 (2026-04-06-rime-mist-v3) and Coda (2026-05-19-rime-coda-tts) to this hear-layer model; because the application decides when an utterance ends, the 105 ms finalisation figure cannot be compared directly with STT products that bundle endpointing.

Details3 sources · 1 X post

Applications choose English or Spanish per request and receive clean or word-for-word output with rewrite-as-it-goes or add-as-it-goes partials.

Rime STT

Rime STT is here, and to celebrate we're giving it away FREE for the next 30 days. Sign up for a free account and snag your API key: https://t.co/s4nfaNDK7V Learn more: https://t.co/t2zicJiirA

https://x.com/rimelabs/status/2112601020698075514
Announcementdocs.rime.airime.aiX
Otter.aiPlatformfunding2026-10-20

Otter.ai closes $320 million in non-equity growth financing from General Catalyst

The sum is more than six times the $50 million Otter had raised through 2021 (PitchBook, via TechCrunch), but it comes from a fund repaid out of sales and marketing returns rather than a priced round, so there is no valuation, and Otter gave no updated ARR, active-user, or churn figures.

Also Use

Details2 sources

The meeting transcription company will use General Catalyst's Customer Value Fund capital for enterprise growth, go-to-market, international expansion, and product innovation.

Otter.ai $320M growth financing

Announcementtechcrunch.com
Mistral AIModel providermodel release2026-10-20

Mistral AI releases Voxtral 3 transcription models with diarisation and end-of-turn detection

Voxtral Transcribe 2 (2026-02-04-mistral-voxtral-transcribe-2, W06) already had a realtime model; version 3 adds native end-of-turn detection, so the STT can make the turn decision that many stacks give to a separate detector, though the 18.4% cpWER and FLEURS lead are Mistral's own results.

Details3 sources · 2 X posts

Small Transcribe 3 covers more than 50 languages with diarisation, context biasing, and native end-of-turn detection, while the Mini variants add a 4B model and a realtime path.

Voxtral Small Transcribe 3, Mini Transcribe 3, and Mini Transcribe Realtime 3

Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵

https://x.com/MistralAI/status/2112574615318876178

Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing. It detects when someone has finished speaking so agents can respond without delay.

https://x.com/MistralAI/status/2112574618586247583
Announcementdocs.mistral.aidocs.mistral.aiXX
HamletEnterpriselaunch2026-10-20net-new

Hamlet opens 10,000+ city-council profiles built on AssemblyAI Universal-3.5 Pro transcripts

AssemblyAI's Universal-3.5 Pro capability (2026-08-19-assemblyai-universal-3-5-pro) now has a civic-information use case, but no council coverage or profile accuracy figure is supplied.

Details2 sources · 1 X post

Residents and journalists can now read Hamlet's profiles of council meetings, where development and rezoning decisions are made.

Hamlet city council profiles · vendor AssemblyAI

City councils decide on real estate development and rezonings in meetings almost nobody watches. Today Hamlet launches 10,000+ profiles, available for residents and journalists, powered by AssemblyAI Universal-3.5 Pro and speaker diarization. Read the full story: https://www.assemblyai.com/customers/hamlet

https://x.com/AssemblyAI/status/2112608442586427821
Announcementmyhamlet.comX
FluidVoiceModel providermodel release2026-10-20net-new

FluidVoice releases Fluid Speech Mini, an open dictation model that keeps transcribing the right speaker when others talk

Underdog's 2026-10-15-underdog-bark-3-0 and FluidVoice's model are two Parakeet fine-tunes released within five days, but their self-reported results use seven versus eight English sets, so 4.94% and 4.54% are not directly comparable. FluidVoice's 10.9% Google SVQ background-speech score against Parakeet Ultra's 14.3% isolates the model's claimed strength to competing voices, while the card reports no equivalent lead under traffic noise.

Details3 sources · 1 X post

FluidVoice published its English dictation model as open CC BY 4.0 weights in Core ML, ONNX, and GGUF builds for iPhone, Mac, Windows, and Linux.

Fluid Speech Mini

Today we're introducing Fluid Speech Mini - the Best Open-Source speech model for on-device Dictation Fluid Speech Mini surpasses Whisper Large, and Parakeet in Accuracy in noisy environments, While being ~10x smaller than Whisper Large, and ~5x smaller than Parakeet. So you can run it for on-device accurate dictation on your MacBook, or even iPhone 11. More info on Fluid Speech Mini below!

https://x.com/FluidVoiceApp/status/2112640613812478133
Announcementhuggingface.cohuggingface.coX
Ant GroupModel providermodel release2026-10-19

Ant Group open-sources Ming-3.0-audio and Ming-3.0-asr for audio understanding and context-aware ASR

The 4.41% average error is an Ant claim, and the audio model's tool requests create an analytical comparison with Alibaba's Qwen Audio Agent 2.0 (2026-09-23-alibaba-qwen-audio-agent-2-0), not a documented relationship.

Also Think

Details6 sources · 2 X posts

Ming-3.0-audio summarises long recordings, answers spoken questions, and generates tool requests, while Ming-3.0-asr-tiny handles accents, dialects, and hotwords without extra training.

Ming-3.0-audio and Ming-3.0-asr

Audio for meetings, interviews & voice-driven services. Ming-3.0-audio-flash summarizes long recordings, answers spoken questions, understands music & ambient sound, and generates tool requests for apps to execute. Ming-3.0-audio-flash offers a more efficient deployment option.

https://x.com/AntLingAGI/status/2112210819366945222

ASR for customer calls, voice search & domain terms. Ming-3.0-asr-tiny handles dialects & accents. Context ASR uses hotwords and prior dialogue without extra training. 4.41% average error across four ASR categories. See the workflow in the demo.

https://x.com/AntLingAGI/status/2112210822483288203
Announcementhuggingface.cohuggingface.cohuggingface.cogithub.comorcarouter.aiXX

Speaktext-to-speech9

ZoomPlatformlaunch2026-10-20

Zoom adds translated meeting audio and receptionist call screening to Workplace and Zoom Phone

Most of these voice features are dated November or December 2026, so only WhatsApp in Zoom Phone is live now; Voice Translator's planned voice cloning aims to keep a speaker's own voice across languages, the same goal as Gradium Voice Localization (2026-10-08-gradium-voice-localization, W41).

Also ConnectUse

Details2 sources

Zoom's meeting and phone products gain translated audio, on-device voice smoothing, and a receptionist that screens and routes calls.

Zoom Workplace Voice Translator and Virtual Agent Receptionist call screening

Announcementnews.zoom.com
Scaled CognitionModel providermodel release2026-10-20net-new

Scaled Cognition launches Voice, a contact-centre TTS model priced from $1 per million characters self-hosted

At $1 per million characters, self-hosting undercuts the $35 to $100 hosted range Akshay Kannan cites, but that comparison excludes GPUs: his estimate is about $13 per million characters at 500 million a month. The preference claim comes from Scaled Cognition's own blind test, while APT customers such as Genesys (2026-09-02-genesys-agentic-virtual-agent) can take the voice from the same vendor.

Details3 sources · 1 X post

Scaled Cognition made its first text-to-speech model generally available for contact-centre calls, either self-hosted in a customer's cloud or hosted by Scaled Cognition.

Scaled Cognition Voice

Today we’re launching Scaled Cognition Voice, the first TTS built for enterprise CX. ⚡ 70 ms time-to-first-audio 🌍 25+ languages and voice cloning 🔒 $1 per 1M characters self-hosted ☁️ $9 per 1M characters hosted by us Enterprise CX teams no longer have to choose between the voice that sounds best and the voice they can afford to put on every call.

https://x.com/ScaledCognition/status/2112562617638072676
Announcementscaledcognition.comsiliconangle.comX
Mistral AIModel providermodel release2026-10-20

Mistral AI releases Voxtral Speech 3 TTS with tagged emotion and pace control and Pro Voice cloning

At $13 per 1M characters, Speech 3 is $3 below Mistral's 23 March Voxtral TTS row (2026-03-23-mistral-voxtral-tts); the vendor's competitor-price comparison does not establish equivalent quality or usage limits.

Details2 sources · 2 X posts

Voxtral Speech 3 brings contextual and tagged delivery control, voice cloning, and professional voice adaptation to more than 35 languages and 50 dialects.

Voxtral Mini TTS 3 (Voxtral Speech 3) and Pro Voice

Voxtral Mini TTS 3 achieves state-of-the-art expressiveness in 35+ languages and 50+ dialects. Implicit instruction-following reads emotion, pace, and emphasis from the conversation. With explicit instruction-following, tags like [emotion=sad][pace=slow] set them line by line.

https://x.com/MistralAI/status/2112574621182775528

With Pro Voice, we’re also rolling out professional voice adaptation. Record up to 30 minutes of audio to replicate your voice with extreme fidelity across every supported language.

https://x.com/MistralAI/status/2112574624034677048
Announcementdocs.mistral.aiXX
KyutaiModel providermodel release2026-10-20

Kyutai shrinks Pocket TTS to 18M, 7M, and 2M-parameter models that keep voice cloning

Kyutai says the 18.2M checkpoint matches the ~100M Pocket TTS it open-sourced in January (2026-01-13-kyutai-pocket-tts) at under a fifth of the size; the 2M models are experimental, so the 103x real-time figure carries no quality claim, and Gradium's same-day Phonon (2026-10-20-gradium-phonon-open-weights) runs on a runtime built on Pocket TTS.

Details2 sources · 2 X posts

Pocket TTS v4.0.0 adds six English checkpoints, from 18M down to experimental 2M variants, plus standalone runtime binaries.

Pocket TTS 4.0.0 small models

Pocket TTS has been getting updates those last few days. While it was ~100M parameters, we release smaller versions, down to 1.9M: an 18.2M model with the quality of the original, a 6.9M model, and a 1.9M model that runs 103x faster than real time on one CPU core. All with voice cloning.

https://x.com/kyutai_labs/status/2112554656459084053

We’ve leveraged the latest open source inference framework by @lmazare and @GradiumAI to speed up generation with our new pocket-tts models on a wide variety of devices. Also excited to see the open source family of pocket-tts based models growing! Go train and contribute your own 🙌

https://x.com/kyutai_labs/status/2112565301816475970
Announcementgithub.comXX
GradiumModel providermodel release2026-10-20

Gradium open-sources Phonon, on-device TTS models under 50M parameters

Phonon turns the 1 April private-beta filing (2026-04-01-gradium-phonon) into downloadable weights; unlike Kyutai's same-day Pocket TTS checkpoints (2026-10-20-kyutai-pocket-tts-small-models), its 31m model uses Apache 2.0, while the 40m variants are CC BY-NC.

Details6 sources · 1 X post

The open release spans an English Apache-licensed model and multilingual CC BY-NC variants, with runtimes for CPUs, GPUs, browsers, and NPUs.

Phonon open weights (Phonon 31m, 40m, and 40m multilingual)

A new frontier for on-device speech. We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date. Phonon-31m is available under Apache 2.0. https://gradium.ai/phonon

https://x.com/GradiumAI/status/2112563128487514512
Announcementgradium.aigithub.comhuggingface.coalphasignal.aix.comX
AmazonModel providermodel release2026-10-20

Amazon upgrades the Polly Generative engine with a new model, 14 voices, and Hindi and Finnish

Polly now pairs the March bidirectional streaming API (2026-03-20-amazon-polly-generative-bidirectional-streaming) with a refreshed model on the same day as Mistral's Voxtral Speech 3 (2026-10-20-mistral-voxtral-speech-3), but AWS publishes no latency figure or benchmark, so the naturalness and accuracy claims remain vendor-stated.

Details1 source

Existing Polly Generative customers get the new model with no API change, and the engine now also runs in AWS GovCloud (US-West) for public-sector and regulated buyers.

Amazon Polly Generative engine

Announcement
GradiumModel providerpartnership2026-10-19

Gradium brings TTS, speech recognition, and translation to Eden AI

The blog describes a 500+ model gateway while the announcement tweet says 1,000+; its documented async jobs are for recordings, not live transcription.

Also Hear

Details1 source · 1 X post

Eden AI users can access Gradium TTS, STT, and speech translation through its AI gateway, with the services pinned to the EU region.

Gradium on Eden AI · vendor Eden AI

Gradium is now on @EdenAIco , the AI gateway with 1000+ AI models behind one API. Our TTS, STT and speech translation run there pinned to the EU region, and every voice in our library reaches first audio in 50ms.

https://x.com/GradiumAI/status/2112152499855704287
AnnouncementX
CekuraPlatformlaunch2026-10-19

Cekura launches a text-to-speech benchmark for 26 models

The 104 agent sentences were tested three times per model in one cloud region, but automated scorers supplied the accuracy and naturalness scores; Cekura also sells voice-agent testing. This extends its Turn Bench (2026-10-13-cekura-turn-bench).

Also Orchestrate

Details3 sources · 1 X post

The public leaderboard compares 26 TTS models across speed, reading accuracy, naturalness, and interruption handling using 30,564 requests.

Text-to-Speech Bench

Today we are launching @cekuraAi's Text-to-Speech Bench. We tested 26 TTS models across 30,564 requests to compare their speed, reading accuracy, naturalness, and interruption handling.

https://x.com/tarush_agarwal_/status/2112208348720570405
Announcementgithub.comruntimewire.comX
Ant GroupModel providermodel release2026-10-19

Ant Group open-sources Ming-3.0-tts in dense 0.5B and 7.9B MoE checkpoints

Artificial Analysis gives Ming-3.0-tts a 93.2% Pronunciation Robustness score, tied with Fish Audio Drama 3 Preview; the 0.5B weights appeared on Hugging Face on 24 September, but Ant's launch announcement came on 19 October.

Details6 sources · 2 X posts

The two checkpoints support zero-shot voice cloning, natural-language voice design, and 14 inline delivery tags across seven languages, with Sichuanese and Cantonese dialect control.

Ming-3.0-tts

Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families. - Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech. - Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue. - Ming-3.0-asr: Tiny 8.6B A2B dialects & context ASR.

https://x.com/AntLingAGI/status/2112210811997540475

TTS for stories, narration & voice assistants. Design voices in text. Direct emotion, pace & emphasis. Add laughter or sighs. 7 languages. - #1 open-weight model in AA Controlled Voice (1069 Elo). - #3 on Provider Voice (1188 Elo) - Ties #1 among all models on Pronunciation Robustness (93.2%).

https://x.com/AntLingAGI/status/2112210816149897641
Announcementhuggingface.cohuggingface.conomorecoding.github.iox.comorcarouter.aiXX

Showavatars, video1

AnamPlatformmodel release2026-10-20

Anam ships Cara-4.1 Express, a faster and more expressive real-time avatar model

Because the shared pipeline held recognition, LLM, and voice constant, the study measures avatar preference rather than end-to-end agent quality; it follows Anam's W42 move to self-hosted Deepgram Flux turn-taking (2026-10-14-anam-deepgram-flux-turn-taking), which addressed the conversational side.

Details4 sources · 1 X post

Anam tested the model blind against HeyGen, Tavus, LemonSlice, and Synthesia with 60 raters and published the evaluation code.

Cara-4.1 Express

Our latest real-time avatar model is live 🚀 Cara-4.1 Express builds on top of cara-4 to make it faster and more expressive. Tested blind against Tavus, HeyGen and LemonSlice, with the code public. Try it: https://t.co/8j4o69NGyT https://t.co/E5QrlPCVtX

https://x.com/Anam__ai/status/2112584252676002208
Announcementanam.aianam.aix.comX

ThinkLLMs, speech-to-speech4

PolyAIModel providerAPI2026-10-20

PolyAI opens Dialog-RSN-1 to self-serve API keys

Because the API speaks the OpenAI Realtime wire format, teams can swap Dialog-RSN-1 in by changing a base URL and key, 82 days after the 30 July early-access launch (2026-07-30-polyai-dialog-rsn-1); the under-300 ms median first-token figure and the quality and safety claims are vendor-reported, with no independent benchmark cited.

Also Hear

Details3 sources · 1 X post

Developers can create a studio.poly.ai account, get an API key, and build through PolyAI's ADK, AWS SageMaker, or the public API.

Dialog-RSN-1

Dialog-RSN-1 is now GA, and anyone can get an API key and build on it today. Comparing real-time voice models? Start with the numbers. Median first-token latency is under 300ms, about the length of a blink. It has the highest quality, audio-understanding, and safety scores of any real-time-capable model, and it’s efficient enough to scale to every conversation you have. It’s audio-native, so turn-taking, speech recognition, function calling, and the response all happen in a single pass, with a separate voice output you control. Get started in three steps: 1. Pick your path: our Agent Development Kit (ADK), AWS SageMaker, or the public API, wherever your stack already lives. 2. Create your account and grab your API key. 3. Clone an example, add your key, and talk to it in your browser. Start now here: https://t.co/tWOVdrwSBB

https://x.com/polyaivoice/status/2112514279182143565
Announcementstudio.poly.aidocs.poly.aiX
HyperPlatformlaunch2026-10-20net-new

Hyper adds OpenAI's GPT-Live-1 full-duplex model to its no-code voice-agent builder

Hyper reaches GPT-Live-1 about six weeks after the model entered OpenAI's API (2026-09-10-openai-gpt-live-1, W37) and after Telnyx (2026-09-30-telnyx-gpt-live-voice-ai-assistants) and Vapi (2026-10-02-vapi-gpt-live-open-beta) in W40, so the no-code tier now gets the duplex model too. Basic-plan access ends on 1 November, and Hyper publishes no HOS reliability figures.

Also OrchestrateConnect

Details3 sources

Businesses building agents on Hyper without code can now run GPT-Live-1 on inbound and outbound calls, WhatsApp, and a website widget, with Hyper's own optimisation layer tuning it for real phone lines.

Hyper with GPT-Live-1 · vendor OpenAI

Announcementhyper.schyper.sc
ElevenLabsModel providermodel release2026-10-20

ElevenLabs launches Eleven Realtime speech-to-speech models that start replying in about half a second

Max beats Mini on Big Bench Audio (98.8% against 75.9%), tool use, and start time, so Mini leads only on Full-Duplex-Bench (95.8% against 94.1%); the OpenAI Realtime-compatible socket puts Eleven Realtime in the same wire format as PolyAI's Dialog-RSN-1 (2026-10-20-polyai-dialog-rsn-1-ga), although Journal du Net's 470 ms and 500 ms latency figures disagree with the blog.

Also HearSpeak

Details5 sources · 1 X post

The two models cover more than 90 languages, the full 17,000-plus voice library, and an OpenAI Realtime-compatible WebSocket for ElevenAgents channels.

Eleven Realtime v1 Mini and v1 Max

Introducing Eleven Realtime, a new class of Speech to Speech models. Voice agents usually make you choose between quick replies or precise answers. Eleven Realtime gives you both. It starts talking in half a second and gets spoken reasoning questions right.

https://x.com/ElevenLabs/status/2112550782063247521
Announcementelevenlabs.ioelevenlabs.iojournaldunet.comthe-decoder.deX
SalesforceModel providermodel release2026-10-19

Salesforce unveils Launa-Omni, a voice model that writes its reply before speaking it

Its 59.69% Audio MultiChallenge score sits just under the top public 60.40%, but Salesforce announced no API or general availability; Slack users get only a simplified version through Slackbot Voice Dictation.

Also HearSpeak

Details1 source

Salesforce's voice model shares a latent space between speech and language, then uses a separate in-house TTS model to speak the response.

Launa-Omni

Announcement

Orchestrateframeworks, evals8

Gnani.aiPlatformlaunch2026-10-21

Gnani.ai adds a Prompt Engine that feeds its voice agents only the scenario they are in

Gnani publishes no token, latency, or cost figures for the claimed savings, and no customer names. Per-scenario prompts with global transitions resemble graph-style conversation flows such as Pipecat Flows, so the change sits in Gnani's hosted Agent Builder rather than a new technique; it closes a ten-day launch series that also produced 2026-10-19-gnani-agent-api and 2026-10-20-gnani-whatsapp-ai-agents.

Details1 source · 1 X post

Gnani's scenario-based prompting capability is built into Agent Builder for complex business conversations.

Prompt Engine

Launch Series : Day 10/10 Introducing scenario-based prompting in Gnani AI’s Agent Builder. The agent gets the instructions for the step it is on, so token usage and latency stay low as the conversation grows. Link in the comments. #VoiceAI #AIAgents #AgentBuilder

https://x.com/GnaniAi/status/2112782180950888863
AnnouncementX
ZoomPlatformlaunch2026-10-20

Zoom adds a camera kiosk channel to Zoom Virtual Agent and a shared Unified Context Layer

Unlike Zoom's 2026-06-22 Agent Architect row (2026-06-22-zoom-agent-architect), this release moves beyond builder controls into a physical retail and service channel; Zoom gives no kiosk availability date, so the rollout cannot yet be compared on deployment timing.

Also SeeUse

Details2 sources

Retail, healthcare, and service-centre shoppers can show products to a virtual agent, while human and AI agents share customer memory across Zoom's CX products.

Zoom Virtual Agent kiosk channel and Unified Context Layer

Announcementnews.zoom.com
neoAIPlatformlaunch2026-10-20net-new

neoAI launches a contact-centre voice agent that completes identity checks and contract changes

The agent extends neoAI Chat and neoAI Work, which neoAI says run at more than 250 companies, and arrives on the same day as wevnal's BOTCHAN AICALL general release (2026-10-20-wevnal-botchan-aicall); neoAI names no customer, price, or speech vendor.

Also Use

Details1 source

The agent handles identity verification, enquiries, and contract changes on customer calls, then writes call records and minutes and arranges follow-up, passing to a person only when needed.

neoAI voice agent

Announcement
Mistral AIModel providerAPI2026-10-20

Mistral AI introduces a Voxtral 3 voice agent SDK built on LiveKit

Building on LiveKit rather than a proprietary runtime makes the pipeline familiar to existing LiveKit Agents deployments, and the separate stages let teams use a non-Mistral LLM; Mistral publishes no end-to-end latency or price for the assembled agent.

Also HearSpeak

Details2 sources · 1 X post

The Voxtral 3 SDK combines realtime transcription, model reasoning, tool calls, and speech output in a streaming pipeline for phone and web agents.

Voxtral 3 voice agent SDK · delivered by LiveKit

Voxtral 3 also introduces a voice agent SDK, built on LiveKit. Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.

https://x.com/MistralAI/status/2112574627381936345
Announcementdocs.mistral.aiX
LiveKitPlatformpricing2026-10-20

LiveKit opens its turn-detection and adaptive-interruption models to self-hosted agents in production, metered per request

Because charges fire per prediction rather than per minute, a call with many 200 ms pauses or 100 ms overlap checks can cost more than a minutes-based estimate; the free hosted path and monthly allowances also make LiveKit Cloud the simpler production route. This follows LiveKit's Turn Detector v1 (2026-06-17-livekit-turn-detector-v1) and Agents 1.6.7 adaptive-interruption work (2026-07-25-livekit-agents-1-6-7-orchestrate).

Details3 sources · 1 X post

The change replaces a monthly cap for self-hosted agents with per-project development allowances and per-request billing on paid plans, while agents deployed to LiveKit Cloud keep both models free.

LiveKit turn detection and adaptive interruption models

Two small models we trained for voice agents: turn detection knows when the user is actually done talking, adaptive interruption knows “uh huh” isn’t an interruption. Free for agents on LiveKit Cloud. Starting today, self-hosted agents can run them in production too: a free development allowance, then metered per request. The old monthly cap is gone. https://docs.livekit.io/agents/logic/turns/

https://x.com/livekit/status/2112639691690279223
Announcementdocs.livekit.iolivekit.comX
NVIDIAPlatformlaunch2026-10-19

NVIDIA's Nemotron Voice Agent 2.3.0 prefills the LLM while the user is still speaking

Streaming-input prefill and 320 ms end-of-turn finalisation are NVIDIA's own latency levers on top of Pipecat 1.12.0, which already trails Pipecat 1.13.0 (2026-10-18-pipecat-1-13-0-orchestrate, 18 October). DGX Spark loses speculative decoding because it regressed with streaming prefill, so the gain is not uniform across NVIDIA's own hardware.

Details1 source

The v2.3.0 blueprint brings Nemotron 3.5 services, speaker labels, consolidated service configuration, and an upgraded Pipecat runtime to its examples.

Nemotron Voice Agent v2.3.0

Announcement
monday.comPlatformlaunch2026-10-19net-new

Harmony turns enterprise sales conversations into voice-agent workflows

Bright Car Finance’s reported 9% lift is the only outcome figure; the platform’s claimed contact rate reaches 70%, with no independent measurement supplied.

Also ThinkConnect

Details2 sources

monday.com built Harmony, a voice-agent platform for enterprise sales teams, on OneAI technology and task-specific small language models.

Harmony · delivered by Twilio

Announcementharmony.ai
Gnani.aiPlatformlaunch2026-10-19

Gnani introduces an API suite for building and managing voice agents

Integration is guided by Gnani's team rather than self-serve, so the API targets enterprise and partner builds; it follows Agentic Evals (2026-10-18-gnani-agentic-evals) and AI Insights (2026-10-17-gnani-ai-insights) in the same launch series.

Also Connect

Details2 sources · 1 X post

Gnani’s API family lets enterprises and platform partners build voice agents, connect workflows, and return conversation data to their systems.

Gnani AI Agent API

Launch Series : Day 8/10 Introducing Gnani AI Agent API. Build and manage voice agents from your own platform, connect existing workflows, and bring conversation data back into your systems.

https://x.com/GnaniAi/status/2112057405232324939
Announcementtipranks.comX

Connecttelephony, channels3

Gnani.aiPlatformlaunch2026-10-20

Gnani launches WhatsApp AI Agents that swap documents and options during a voice call

Day 9 of Gnani's ten-day launch series extends the previous day's API filing (2026-10-19-gnani-agent-api), but the named examples do not identify the lending, banking, or hospitality customers.

Also Orchestrate

Details1 source · 1 X post

Gnani's agents let callers exchange messages, images, and documents in the same WhatsApp thread while a voice call continues.

WhatsApp AI Agents

Launch Series : Day 9/10 A phone call usually gives you words. Nothing to look at, nothing to scroll back to, nothing to tap on. But some moments need more than talking, a menu to pick from, a confirmation to glance at, a step that’s easier shown than said. Our AI voice agent on WhatsApp handles exactly that: it carries the call like any other, and the instant something’s easier to show than say, it lands straight in the same chat, no separate app, no “check your email.” The call and the chat aren’t two channels. They’re one conversation, working the way a conversation should. #VoiceAI #ConversationalAI #WhatsAppBusiness #CustomerExperience #Gnani #AIAgents

https://x.com/GnaniAi/status/2112419793227694306
AnnouncementX
call4mePlatformlaunch2026-10-20

call4me adds inbound answering, a REST API, and x402 payments for agent phone calls

Inbound answering moves call4me from outbound errands (2026-09-30-call4me-agent-phone-calls, W40) into the business-line job that BotPhone MCP (2026-10-20-botphone-mcp-agent-phone-line) launched on Product Hunt the same day, and x402 lets an agent buy a call in USDC on Base without a human creating an account. The tripled user base is maker-reported with no base figure.

Also Use

Details3 sources

A call4me number can now answer incoming calls from saved standing instructions, and developers and agents can place calls through a REST API or pay per call with x402 instead of using MCP.

call4me inbound answering, REST API, and x402

Announcementcall4.mecall4.me
3CXPlatformlaunch2026-10-19net-new

3CX adds call scripts that run OpenAI Realtime, Gemini Live, or Grok voice agents inside the PBX

The scripts keep provider API keys and the realtime model as protected PBX parameters, while a failed session falls back to a normal 3CX destination; the blog names no model versions or prices.

Also Orchestrate

Details1 source

The Call Script store covers OpenAI Realtime, Google Gemini Live, and xAI Grok real-time voice, with PBX routing, transfer tools, and fallback destinations.

AI voice agent Call Scripts · vendor OpenAI, Google, xAI

Announcement

Useagents in market11

Channel CorporationPlatformlaunch2026-10-21net-new

Channel Corporation's Channel Talk launches Voice ALF, a phone agent that answers enquiries and calls customers first

The in-house turn-detection and Korean TTS make Channel Talk a chat-first CX platform that is building its own speech layer rather than buying it. The 87% within 1.5 seconds result comes from Channel Talk's own café orders, while Lincle's 39% and 42% reductions are vendor-reported; plan availability on 16 October preceded the 21 October announcement.

Also SpeakOrchestrate

Details5 sources

Voice ALF extends Channel Talk's chat-based ALF agent to inbound and outbound customer counselling.

Voice ALF

Announcementdocs.channel.ioinews24.comdigitaltoday.co.krasiae.co.kr
YelpPlatformlaunch2026-10-20

Yelp's Check for Me sends an AI voice agent to call restaurants about wait times

The caller-side flow complements Yelp Host's restaurant-side answering (2026-07-28-yelp-host-opentable-food-ordering, W31, and 2026-09-10-yelp-host-hatch-gpt-live-1, W37), while the coming-months rollout and eligible-page limits leave Check for Me's coverage unknown.

Also Connect

Details3 sources

Yelp's app can ask an AI voice agent to call an eligible US restaurant and text the user its current wait time.

Check for Me

Announcementblog.yelp.comtechcrunch.com
wevnalPlatformlaunch2026-10-20

wevnal moves its BOTCHAN AICALL phone agent from beta to general release

wevnal says AI first-line handling of abusive callers supports Japan's amended labour law on customer harassment, in force since 1 October 2026. Vendor-reported results include 100% call completion and about 3.1 times the first month's CVR for Orico Car Life after four months, 80% automation of NR's cancellation calls with retention close to human agents, and an answer rate above 96% at Locabo Works while operator seats were nearly halved. The company targets 100 customer companies by 2027 and 500 by 2030; the wider BOTCHAN platform is used by more than 800 companies.

Also Orchestrate

Details2 sources · 1 X post

wevnal's phone agent automates enquiries, orders, cancellations and changes, bookings, and pricing or how-to questions 24 hours a day, 365 days a year.

BOTCHAN AICALL

電話AIエージェント「BOTCHAN AICALL」を本日正式リリースしました📣 問い合わせ・注文・解約・変更などの電話対応を、熟練オペレーター並みの品質で24時間365日自動化します。 #BOTCHANAICALL https://t.co/QeI6its4D6

https://x.com/wevnalinc/status/2112393460582166950
Announcementbotchan.ioX
SpeakONPlatformlaunch2026-10-20net-new

SpeakON turns spoken intent into approved actions with SpeakON Agent

SpeakON is US-only and gives no pricing or adoption figures; where Willow's Wally (2026-10-19-willow-wally-scribe-agent) drafts by voice on the desktop, SpeakON puts dictation behind a physical button on the iPhone (comparison only).

Also Hear

Details1 source

The iPhone product pairs a magnetic AI button with an AI Voice Keyboard that cleans dictated text and turns recordings into lists or notes in the app already in use.

SpeakON Agent

Announcement
SierraPlatformpartnership2026-10-20

Sierra and Stripe let agents take card payments inside the call, with no transfer or payment link

The reported conversion covers only calls that began a payment, so it cannot establish overall payment completion; the integration gives Sierra's W41 Personal Agent Protocol filings (2026-10-06-sierra-meta-personal-agent-protocol, 2026-10-09-sierra-personal-agent-protocol-poppy-draft) a concrete transaction path.

Also Connect

Details2 sources · 1 X post

Cedar's Kora is an early customer, with Sierra reporting successful payment on 71.3% of its payment-starting calls.

Sierra payments with Stripe · delivered by Stripe

We've partnered with @stripe so customers can get answers and pay in the same conversation, with no transfer and no payment link. @CedarNY is one of the first to take phone payments through the integration. When a patient calls Cedar's agent Kora about a bill, they can ask about a charge, pay securely on their keypad, and get confirmation without leaving the call — and 71.3% of those calls end in a successful payment. "Medical bills create a moment of real anxiety for patients, and historically, getting a question answered and actually paying the bill have been two separate, disconnected steps. Working with Sierra and Stripe, we've closed that gap. A patient can call confused about a charge and leave the conversation with clarity and a resolved balance, in the same call." — Dugan Winkie, Chief Strategy Officer, Cedar

https://x.com/SierraPlatform/status/2112604880171131197
Announcementunite.aiX
InvocaPlatformlaunch2026-10-20

Invoca launches healthcare voice and messaging agents that book appointments in Epic and athenahealth

University Hospitals' first-night result is a single customer outcome, so it does not establish booking lift across the five integrations; the healthcare release extends the same vertical playbook as Invoca's home-services agents (2026-10-06-invoca-home-services-ai-agents).

Also Orchestrate

Details3 sources · 1 X post

Invoca's Data + AI healthcare solution pairs HIPAA-compliant agents with scheduling and CRM write-back, and is available now.

Invoca Data + AI healthcare solution

We’re excited to launch Invoca’s new data + AI healthcare solution. By integrating data with our new AI agent, we’re able to provide the healthcare industry with new opportunities and benefits. 🧵 below

https://x.com/Invoca/status/2112562000329064918
Announcementprnewswire.cominvoca.comX
BotPhonePlatformlaunch2026-10-20net-new

BotPhone launches BotPhone MCP, a business phone line that personal agents such as Muse and Grok Bot can run

BotPhone sells the inbound receptionist side that ClawCall (2026-10-05-clawcall-chatgpt-plugin, W41) only offers in its $14.99 tier, and it competes head-on with call4me's same-day inbound answering (2026-10-20-call4me-inbound-rest-api-x402); at $0.20 a minute it undercuts call4me's $0.25. Earlier availability to Muse users since at least 24 September means the Product Hunt date is the formal launch, not the first call.

Also Connect

Details3 sources · 1 X post

Users of MCP-capable assistants get a real number whose voice agent answers and places calls, while their own agent sets it up, sends it on calls, and reads back a summary, transcript, and next step.

BotPhone MCP

Just shipped Calendly booking for the BotPhone MCP. Your @Muse, @bot, and other personal agents can now screen a caller, pick the right meeting type, and book them straight onto your Calendly while they’re still on the phone. Connect it in one click from the dashboard.

https://x.com/garrickbridges_/status/2110385150240661697
Announcementproducthunt.combotphone.telX
Batteries PlusEnterprisepartnership2026-10-20net-new

Batteries Plus deploys Synthflow voice agents on Zoom Phone

The 700-location rollout makes the marketplace listing materially different from a sandbox integration, but both the under-half-second response and five-week rollout are supplier or customer-reported rather than independent benchmarks.

Also Connect

Details5 sources

Batteries Plus is the first at-scale customer for Synthflow's Zoom Phone deployment, spanning more than 700 US locations.

Synthflow voice agents on Zoom Phone · vendor Synthflow · delivered by Zoom

Announcementfinanznachrichten.dedocs.synthflow.aidocs.synthflow.aisynthflow.ai
WillowPlatformlaunch2026-10-19

Willow adds Wally, a Scribe agent that drafts and rewrites text through a spoken back-and-forth and learns from the user's edits

Two weeks after Willow Knowledge (2026-10-06-willow-knowledge, W41), that imported context now feeds a multi-turn agent instead of one-shot dictation clean-up, moving Willow from transcription towards voice-driven drafting. The three-edit learning figure comes from the team's own Product Hunt reply, not a measured result.

Also Hear

Details2 sources

Willow dictation users can now talk or type to a writing agent that holds a running conversation over a draft, rewrites text highlighted in another app, and pulls in the personal context imported through Willow Knowledge.

Wally

Announcementwillowvoice.com
Hunar.aiPlatformfunding2026-10-19net-new

Hunar.ai raises Rs 39.3 crore for frontline hiring agents

Entrackr’s estimate of about Rs 180 crore values Hunar.ai at roughly 100 times its FY25 revenue of Rs 1.70 crore, so investors are pricing frontline voice hiring well ahead of revenue; the 700–900 ms response time is the company’s own claim.

Details2 sources

Blume Ventures led the Series A for multilingual voice agents that screen, onboard, train, and engage frontline workers.

Hunar Voice

Announcementblume.vc
Deutsche TelekomEnterpriselaunch2026-10-19

Deutsche Telekom launches Voice AI Notes Mobile call summaries on AudioCodes VoiceAI Connect

This is a new mobile-business service rather than a revival of Deutsche Telekom's W10 Magenta AI Call Assistant (2026-03-02-deutsche-telekom-magenta-ai-call-assistant), which used ElevenLabs; AudioCodes gives no pricing, customer numbers, or availability date.

Also HearConnect

Details1 source · 1 X post

The service targets small and medium-sized businesses and uses AudioCodes VoiceAI Connect with Mistral AI to process mobile-network calls.

Voice AI Notes Mobile · vendor AudioCodes (VoiceAI Connect), Mistral AI

Deutsche Telekom Selects AudioCodes VoiceAI Connect to Power AI-Driven Mobile Business Services. Business calls can now do more than just connect people. With Voice AI Notes Mobile, Deutsche Telekom - a leading German telecommunications provider - is helping business customers automatically record, transcribe and summarize calls, giving employees more time to focus on their customers and core business activities. Behind the service, AudioCodes VoiceAI Connect connects the mobile network with AI services while giving Deutsche Telekom the flexibility to evolve its Voice AI capabilities over time. And this is just the beginning! Get all the details >>> https://t.co/NvBDDo9yoK

https://x.com/AudioCodes/status/2112157232213201048
AnnouncementX