Hearspeech-to-text22
SpeechFast runs hold-to-talk dictation for the Mac entirely on the device, with a choice of Parakeet, Cohere Transcribe, or Apple engines
Cohere Transcribe shows up here as a selectable engine in a paid consumer app, next to Parakeet, which T-Scriptor (2026-10-13-t-scriptor-on-device-parakeet) also uses this week; despite the 'Local Whisper' tagline, none of the three engines on the site is Whisper.
Details2 sources
A one-time-purchase Mac app types cleaned-up speech at the cursor in any app, with transcription on the Neural Engine and filler removal by Apple Intelligence, for people who cannot send audio to a server.
SpeechFast
Pipecat 1.13.0 adds Pulse 2.0 and updates transcription defaults
The AssemblyAI default adopts the model filed as 2026-09-29-assemblyai-universal-3-6-pro-realtime; teams relying on earlier models need to pin them, while Pulse 2.0 ignores the prior endpointing setting.
Details3 sources
SmallestSTTService supports English-only Pulse 2.0 with built-in end-of-turn detection and optional emotion and gender labels on final transcripts.
pipecat-ai 1.13.0
fal hosts Kotoba speech APIs for transcription, speech translation, and Japanese speech generation
Kotoba's Japanese-first speech stack, first opened as an alpha API on 28 May 2026 (2026-05-28-kotoba-api-sdk-alpha), now bills per use through an existing fal key; caveat: third-party trackers saw the model pages from 9 October, before the 18 October announcement.
Also Speak
Details5 sources · 2 X posts
Developers can call Kotoba batch and streaming transcription, real-time speech translation, and Japanese text-to-speech through fal endpoints and the kotoba-sdk.
Kotoba speech APIs on fal · vendor Kotoba
Kotoba's speech APIs are now live on fal. - Translates speech in real time with text and audio output - Transcribes live audio as it streams - Generates speech with preset voices - Transcribes recordings with optional word timestamps
https://x.com/fal/status/2111971103765639510
Kotoba is now on @fal! 🚀 From speech recognition and natural Japanese speech synthesis to real-time speech translation, we’re excited to bring our voice AI to more developers. We can’t wait to see what you build! #Kotoba #fal #speechAPI
https://x.com/kotoba_tech/status/2111984917408297155
EarnotePlatformlaunch2026-10-17net-new
Earnote transcribes lectures with Whisper on the Mac or iPhone and writes notes, flashcards, and a study sheet without a cloud account
T-Scriptor (2026-10-13-t-scriptor-on-device-parakeet) filed the same lecture job four days earlier on Parakeet; Earnote stays on Whisper but, on iPhone 15 Pro and later, writes the note on the phone too, so a whole lecture can be processed with no network (analytical).
Also Use
Details4 sources
The MIT-licensed Mac app and its new iPhone and iPad version record lectures, meetings, or calls, transcribe them on the device, and turn the transcript into a summary, tasks, flashcards, and a PDF study sheet, with the note itself also written on-device on iPhone 15 Pro and later.
Earnote 1.0 (Mac, iPhone, iPad)
ScriebPlatformlaunch2026-10-16net-new
Scrieb 1.0.0 strips speaker echo from offline call recordings on Mac and Windows and detects the spoken language from speech
Earnote (2026-10-17-earnote-on-device-lecture-notes), filed the next day, rules out Windows because it depends on Apple's speech and AI frameworks, and SaidTrace (2026-10-15-saidtrace-mac-voice-memo-archive) is Mac-only, so Scrieb is the one local-Whisper note-taker this week with a native Windows on Arm build (analytical); its no-upload pitch to clinicians rests on the vendor's own data-flow page at scrieb.com/security.
Also Use
Details3 sources
The desktop note-taker records calls through system audio without a bot, or imports files, and transcribes, labels speakers, summarises, and answers questions on the user's own computer, pitched at lawyers, therapists, clinicians, and researchers.
Scrieb 1.0.0
MayblocModel providermodel release2026-10-16net-new
Maybloc launches Calico, a speech-to-text model that keeps overlapping speech and separates up to 25 speakers
Crosstalk in meetings is where diarisation errors concentrate, and Calico's 17.7 cpWER on NOTSOFAR against 31.2 for AssemblyAI Universal-3.5 Pro comes from Maybloc's own run with no third-party check yet. Table Canon (2026-10-14-table-canon-ttrpg-session-memory) assembles the same job from WhisperX and pyannote 3.1, and VocaScript (2026-10-09-vocascript-long-form-transcription, W41) sells long-form speaker labels in the browser (analytical; neither uses Maybloc).
Details5 sources
Maybloc's REST API and dashboard transcribe uploaded recordings of up to three hours with its own Calico model, returning per-speaker turns, overlap spans, and word timings in 45 languages for $0.006 a minute.
Calico (calico-1-flash)
UnderdogModel providermodel release2026-10-15net-new
Underdog open-sources Bark 3.0, a 163 MB on-device English speech-to-text model fine-tuned from Parakeet
Bark's 4.94% average and 7.82% AMI result are self-run with the leaderboard's code rather than public-board entries, and its 0.004-point edge over Parakeet at three decimals is a size win rather than an architecture change: the 163 MB model is a Parakeet fine-tune. FluidVoice's 2026-10-20-fluidvoice-fluid-speech-mini uses eight rather than seven sets, so the two percentages are not directly comparable.
Details2 sources · 1 X post
Underdog released four open CC BY 4.0 English speech-recognition checkpoints for on-device dictation and meeting notes, with the 163 MB Bark 3.0 running locally in its Mac app.
Underdog Bark 3.0
Today we’re releasing Underdog Bark 3.0, the #1 frontier speech model
It's a tiny 163 MB download as accurate as NVIDIA’s Parakeet (15x bigger). It beats OpenAI's Whisper large-v3 (57x bigger)
It's fast- transcribes 1 min of audio in <0.25s on a Macbook. On recorded meetings (AMI), Bark 3.0 gets 7.82% of words wrong. Parakeet gets 9.42%. OpenAI's Whisper large-v3 gets 13.63% error rate
Bark runs on both your Laptop & Phone. It’s free and open source under CC-BY 4.0
Try private on-device meeting notes & dictation, powered by Bark, in Underdog today: http://underdog.ai/bark
https://x.com/UnderdogAI/status/2112733007484518470
SaidTrace transcribes iPhone, Apple Watch, and recorder audio on the Mac and links every AI answer back to the recording
SaidTrace prices the archive at $3 a month against the pendant recorders on its own comparison table, such as Plaud NotePin S at $179 plus $29.99 a month, by using an iPhone or recorder the user already owns. Its MCP link to coding assistants is only partly tested by the maker's 3 October checks.
Also Use
Details2 sources
SaidTrace imports voice memos from an iPhone, an Apple Watch, or a USB recorder into one archive on the Mac, transcribes them locally with Whisper, and lets the user pick the AI that writes notes and answers questions.
SaidTrace 0.1.5 for Mac · vendor OpenAI Whisper (run locally)
IFM releases Jais-ASR, an open Arabic speech recogniser for dialects from 19 countries
A second open Arabic recogniser from Abu Dhabi in nine days, after TII's Falcon-ASR (2026-10-06-tii-falcon-asr, W41) for Emirati and Standard Arabic; this one claims dialects from 19 countries and code-switching into four European languages under Apache 2.0. The leaderboard lead over Speechmatics and ElevenLabs is IFM's own measurement.
Details6 sources · 1 X post
MBZUAI's Institute of Foundation Models and G42's Inception42 published Jais Speech under Apache 2.0 with the Jais 3 language models, transcribing Modern Standard Arabic and regional dialects, including mid-sentence switches into English, French, Spanish, and Italian.
Jais-ASR (Jais Speech)
GladiaModel providerpricing2026-10-15
Gladia cuts async transcription by up to 70%, with Solaria-3 from €0.16 an hour
At €0.16 to €0.12 an hour, Gladia's EU-hosted batch STT is now priced near the cheapest US APIs, such as AssemblyAI Universal-3.6 Pro at about $0.30 an hour on OpenRouter, which puts pressure on per-hour pricing for European call analytics.
Details2 sources · 1 X post
Gladia's pre-recorded transcription now starts at €0.16 an hour on Solaria-3 with no commitment, credits that never expire, and up to 20% bonus credit on large top-ups; the new rates apply to existing users automatically.
Gladia async transcription (Solaria-3)
Gladia async transcription is up to 70% cheaper!
- €0.16/hr with Solaria-3
- No commitments; credits never expire
- Up to 20% bonus credits on €1,000+ top-ups
- Same accuracy & security. Auto-applied for existing users.
https://x.com/gladia_io/status/2110702489578480091
AmazonPlatformlaunch2026-10-15
Amazon's Bee adds a wrist-worn Frontier recorder and a phone-mounted Eclipse that also captures calls, both linked to Alexa+
Alexa+ becomes the read-back interface for an always-on recorder, a consumer-memory use beside the eDreams travel booking skill (2026-10-07-edreams-odigeo-alexa-plus-travel) and the India launch (2026-09-16-amazon-alexa-plus-india); the call-capture claim covers only Eclipse, which had no ship date on 15 October.
Also Use
Details2 sources · 3 X posts
Two personal AI devices listen to the wearer's conversations and turn them into summaries, to-dos, and reminders in the Bee mobile and new desktop apps, and Alexa+ on Echo can now read those notes back on request.
Bee Frontier and Bee Eclipse
Table Canon turns tabletop game-night recordings into diarised recaps and a campaign memory game masters can question
Every speech component is an open model named on the launch thread, so what Table Canon sells is memory across sessions and a spelling glossary, the same split T-Scriptor (2026-10-13-t-scriptor-on-device-parakeet) makes between free transcripts and paid notes (analytical). Diarisation of four to six people talking over each other at a table is harder than a two-party call, and no accuracy figures are given.
Also Use
Details2 sources
Game masters upload a session recording and Table Canon writes the recap, outline, and dossiers on people and places, keeps a searchable record across sessions, and suggests prep from threads still open.
Table Canon · vendor WhisperX (Whisper large-v3), pyannote 3.1, and SpeechBrain ECAPA
Speechmatics launches Oak, a multilingual medical speech-to-text model
Ambient scribes that sit in real multilingual consults get a medical-native model that switches languages without a pre-selected pack — Speechmatics' own table puts Oak at 91% medical-term recall across 11 languages while the other eight providers average 74%, so missed drug names and diagnoses shrink on the vendor's test set even if WER looked fine.
Details3 sources · 1 X post
Oak is a medical-native STT model on the Speechmatics API that follows language switches mid-consultation across more than 55 languages, tracks speakers, and is tuned to keep drug names and other clinical terms.
Oak (medical multilingual STT)
Tether's QVAC releases TranscribePsy, a 254M-parameter on-device ASR model
This is W42's third on-device ASR ship after Canonical's Myna preview (2026-10-12-canonical-myna-tech-preview) and T-Scriptor's Parakeet move (2026-10-13-t-scriptor-on-device-parakeet); the 'research and educational purposes' wording beside an Apache 2.0 licence needs checking before commercial use.
Details3 sources · 2 X posts
Tether Data's QVAC team released open-weight multilingual speech-to-text checkpoints that run offline on phones, laptops, smart glasses, and robots, with GGUF and Core ML builds.
QVAC TranscribePsy
We just released QVAC TranscribePsy, a speech-to-text model small enough to run inside a phone app.
It covers 25 European languages on a laptop and 8 on a phone, and on an iPhone 17 Pro Max's Neural Engine the 8-language build transcribes a 10-minute recording in 1.65 seconds.
Glasses, robots and phones are moving to voice, and each of them first has to write down what you said. Today that step mostly happens on a server.
TranscribePsy does it on the phone or laptop, so dictation and meeting notes keep working with no signal, and a private conversation can be transcribed without the recording ever leaving the room.
Run it on our qvac-fabric-speech.cpp's parakeet engine, which optimizes for efficient execution and delivers the highest performance.
Get the models: https://huggingface.co/collections/qvac/transcribepsy
https://x.com/qvac/status/2110369471802290206
https://x.com/tether/status/2110344819411189856
NatWest tunes NVIDIA Nemotron ASR to UK accents for Spending Insights
The release names the speech stack behind Spending Insights, first filed on 25 September (2026-09-25-natwest-spending-insights-voice); the 20% reduction is NatWest's own test on its colleague accent set, with no public benchmark.
Also Use
Details1 source
The bank adapted NVIDIA's open Nemotron speech recognition with colleague recordings and anonymised contact-centre transcripts for its voice and text Spending Insights tool, now available to 75% of Royal Bank of Scotland customers.
Spending Insights (Nemotron ASR) · vendor NVIDIA
AnamPlatformpartnership2026-10-14
Anam moves its avatars' turn-taking to self-hosted Deepgram Flux and halves unwanted barge-ins
Anam added Deepgram as one selectable STT in September (2026-09-09-anam-deepgram-stt-provider); Flux is now the default turn-taking layer for every turnkey avatar, English only for now. The ~50% barge-in cut comes from a vendor customer story with no published test set, unlike the shared runs in Cekura's Turn Bench (2026-10-13-cekura-turn-bench).
Also Show
Details1 source · 1 X post
Anam replaced its open-source speech-to-text and turn-detection stack with Deepgram Flux, self-hosted on its own Kubernetes cluster, for all turnkey users of its real-time interactive avatars.
Anam turnkey avatars on Deepgram Flux · vendor Deepgram
T-Scriptor moves lecture transcription onto the Mac and iPhone with Parakeet and writes the notes after class
The transcript stays free and the paid tier is the lecture notes, which matches OpenWhispr for iOS (2026-08-05-openwhispr-ios, W32), also on Parakeet with a free on-device tier (analytical). The 13 October build is what made the 'audio never leaves your Mac' claim true; the September builds sent audio to the cloud.
Also Use
Details3 sources
Graduate student Bhuvanesh Balaji's lecture recorder now transcribes on the device in 25 European languages, then writes structured notes with key concepts, due dates, and the papers a lecturer mentions.
T-Scriptor 1.0.7 for Mac and iPhone · vendor NVIDIA Parakeet
GladiaModel providerlaunch2026-10-13
Gladia releases an open-source MCP server that lets coding assistants run its transcription jobs
Coding agents get diarised transcripts and summaries but no live audio, since streaming stays out of scope; it is Gladia's first filing since OVH Groupe completed its acquisition (2026-07-31-gladia-ovh, W31), and Bolna took the same MCP route for voice-agent management in July (2026-07-19-bolna-mcp-server-cli, analytical).
Also Orchestrate
Details3 sources · 1 X post
The MIT-licensed `@gladiaio/mcp` package connects Cursor, Claude, Codex, and other MCP hosts to a Gladia account so they can create and manage pre-recorded transcription jobs from a prompt.
Gladia MCP
✨ Meet Gladia MCP: speech-to-text inside your AI agents.
- Transcribe any recording into a transcript or summary with one prompt
- Check transcription jobs without opening a dashboard
Open-source. Works with Cursor, Claude and Codex.
👉 https://github.com/gladiaio/gladia-mcp
https://x.com/gladia_io/status/2109977450533036098
FUTOModel providermodel release2026-10-13net-new
FUTO releases ASR4ALL, streaming speech recognition models sized for CPUs and phones
At 28M to 97M parameters with CPU and streaming targets, ASR4ALL is smaller than QVAC TranscribePsy (254M, 2026-10-14-qvac-transcribepsy), which shipped the next day. FUTO publishes speed claims but no word error rates, so its accuracy against those peers is unverified.
Details3 sources
FUTO publishes small (28M), medium (56M), and large (97M) streaming ASR models with an add-on for punctuation and error correction, a tiny voice activity detector, and 'plus' variants with speaker segmentation and speaker-embedding heads, under a commercial licence with ethical carve-outs.
ASR4ALL
ElevenLabs adds turn-prediction commits to realtime Scribe and audio effects to TTS
End-of-turn detection now arrives inside the STT stream, a job agent frameworks have handled with separate turn detectors; the stereo effects are refused in the μ-law and A-law formats phone lines use, so they serve app and media output more than telephony.
Also SpeakOrchestrate
Details1 source
The 12 October changelog lets the realtime Speech to Text WebSocket commit a segment once the model predicts the speaker has finished, and lets Text to Speech and Text to Dialogue requests apply filters, distance, reverb, background noise, and panning.
Scribe realtime turn prediction; TTS audio effects
Canonical opens Myna local speech-to-text dictation to Ubuntu users as a technology preview
Ubuntu becomes the first major Linux distribution in this corpus to ship first-party offline dictation on Parakeet or Whisper, an OS-vendor counterpart to the W41 Mac dictation app Halo 2.0 (2026-10-11-halo-mac-dictation-notch); Canonical still calls it a rough-edged preview and rules out voice commands.
Details8 sources · 1 X post
Ubuntu Desktop users on 24.04 LTS, 26.04 LTS, and 26.10 can install Myna from a PPA and dictate into whichever app has focus, with recognition running on the machine and no audio uploaded or stored.
Myna · vendor Parakeet, Whisper, and FunASR (local engines)
Ubuntu 26.10 Stonking Stingray has emerged with agility and grace. “Earlier this year, we set out a roadmap for AI on Ubuntu. Stonking Stingray delivers the first milestone through the integration of Myna: a speech-to-text agent that utilizes on-device LLMs.” - Jon Seager, VP of Ubuntu Engineering at Canonical
https://x.com/Canonical/status/2110710303260332424
AmbiPlatformfunding2026-10-12net-new
Mistlabs raises $26M Pre-A led by Luminous Ventures and Huaqin and launches Ambi, an always-on listening assistant for Apple devices
Ambi moves the meeting-notetaker job from after the call to during it (Action Spotter and Routines act while people are still talking), the opposite timing to Amazon's Bee recorders that summarise afterwards (2026-10-15-amazon-bee-frontier-eclipse) and Otter's post-call model funded the following week (2026-10-20-otter-ai-320m-cvf-financing); its consent answer is still 'tell people yourself', and it is hiring to build its own speech recognition rather than naming an STT vendor.
Also Use
Details3 sources · 2 X posts
Ambi listens during meetings and calls on iPhone, Mac, and Apple Watch and acts while the conversation is still running, surfacing suggestions, fact-checking, and drafting follow-up work instead of waiting for a post-call summary; the round funds Mistlabs' own AI wearable.
Ambi (ambient intelligence for iPhone, Mac, and Apple Watch)
Speaktext-to-speech13
VoicedByMe clones a user's voice from 20 seconds of speech on the computer and reads documents, screen text, and coding-agent replies aloud in it
SpeechFast (2026-10-18-speechfast-local-dictation-mac) and T-Scriptor (2026-10-13-t-scriptor-on-device-parakeet) brought on-device speech recognition to consumer desktops this week; VoicedByMe is the speech-output counterpart, with a local OpenAI-compatible endpoint other apps can call (analytical). The site names neither the TTS model nor the company behind it, so its provenance is unverified.
Details2 sources
A one-time-purchase desktop app for Windows and Apple-silicon Macs runs voice cloning and text-to-speech on the device, for reading selections, making audiobooks and voice-note clips, and giving coding agents a spoken voice.
VoicedByMe
Speechify marks Simba 3.5 as its recommended multilingual TTS model
The 21 November routing change affects existing `simba-multilingual` requests, so clients using that ID have a dated transition to account for.
Details1 source
Simba 3.5 supports 48 languages under one model ID, with pronunciation lists available in IPA or alias form for all but English.
Simba 3.5
Pipecat 1.13.0 adds Rime WebSocket v1 and new TTS models
These hooks follow Google Gemini TTS (2026-09-23-google-gemini-3-8-flash-tts), Sarvam Bulbul (2026-10-13-sarvam-bulbul-v4-flash), and ElevenLabs v4 (2026-09-28-elevenlabs-v4); Speechify users must replace the two removed model names.
Details3 sources
Builders can choose RimeV1TTSService, Gemini 3.8 Flash TTS variants, and Sarvam Bulbul v4 Flash in Pipecat.
pipecat-ai 1.13.0
OpenAIModel providerAPI2026-10-16
OpenAI opens prompt-to-voice to all API customers and adds nine natural GPT-Live voices
API builders get native-language Live voices in five of the six languages that reached ChatGPT a day earlier with GPT-Live-1.1 (2026-10-15-openai-gpt-live-1-1), Mexican Spanish being the one missing, while the 1.1 model itself has no API date. Gradium Voice Design (2026-09-07-gradium-voice-design, W37) and Sierra Persona Studio (2026-10-06-sierra-persona-studio, W41) sell comparable describe-or-record voice creation outside OpenAI's stack (analytical).
Also Think
Details3 sources · 1 X post
API developers can now create a saved GPT-Live voice from a text description, free of charge, and pick from nine new preset voices in German, French, Hindi, Japanese, and Korean.
GPT-Live prompt-to-voice and natural voices
BHASHINIModel providermodel release2026-10-16
BHASHINI and IISc release text-to-speech models for Sourashtra, a minority language of Tamil Nadu
A national programme now ships voices for a language with 248,000 speakers by 2011 Census count, and the reusable piece is the 28-hour two-script corpus; this week's commercial Indian TTS launches cover 11 languages (2026-10-13-sarvam-bulbul-v4-flash) and ten (2026-10-14-gnani-timbre-v3), none of them Sourashtra, and BHASHINI's open VoicERA stack (2026-02-18-bhashini-voicera, W08) is the route such voices would take to phone services (analytical).
Details4 sources
Four VITS voices (one male and one female speaker, each for Tamil-script and Sourashtra-script text) for a language with about 248,000 speakers, released with a 28-hour open speech corpus on BHASHINI's AIKOSH platform.
Sourashtra TTS models
Smallest AI's Pulse STT and Lightning TTS plug into Agora's Conversational AI Engine
Agora's engine has mostly been fed by the large speech vendors, and a single vendor flag for both STT and TTS lets Agora builders try Smallest without running their own pipeline; it follows Smallest's Tenstorrent hardware deal (2026-10-02-smallest-tenstorrent-lightning-v2) as a second distribution route this month.
Also Hear
Details5 sources · 1 X post
Agora voice agents can now name Smallest AI as both the speech-to-text and text-to-speech provider in the agent's join config, with Agora running the real-time conversation.
Pulse and Lightning on Agora Conversational AI Engine · vendor Agora
Smallest AI is now integrated with @AgoraIO's Conversational AI Engine.
Build real-time voice agents with Pulse for speech-to-text and Lightning for text-to-speech. Configure both in your agent’s join config, and let @AgoraIO handle the real-time conversation.
https://x.com/smallest_AI/status/2110724392829518147
Maya Research announces Maya-2 Indic TTS as an open-weight release
Maya names Sarvam's Bulbul v3 as a model it beats a day after Sarvam shipped Bulbul V4 Flash (2026-10-13-sarvam-bulbul-v4-flash), and no Maya-2 weights were on Hugging Face at capture on 15 October, so the open-weight claim is still unverified.
Details3 sources · 1 X post
Maya Research says its Maya 2 text-to-speech family for Indian languages will be released as open weights and free for builders, beside a hosted streaming API with Native, Global, and Flash variants.
Maya 2 (Native, Global, Flash)
Gnani.aiModel providermodel release2026-10-14
Gnani AI ships Timbre v3.0, an Indian-language TTS model directed by inline audio tags
Timbre v3.0 is the second Indian-language TTS launch in two days after Sarvam Bulbul V4 Flash (2026-10-13-sarvam-bulbul-v4-flash), and its emotion tags are live while Bulbul V4 Pro's are still undated; the latency claims are Gnani's own, with no third-party test. Gnani's previous voice filing was Vachana STT and TTS in February (2026-02-19-gnani-vachana-stt-tts, W08).
Details4 sources · 1 X post
The model adds 13 inline tags for laughs, sighs, whispers, and pauses to 13 voices that each speak ten Indian languages, on the Gnani API and playground.
Timbre v3.0
Launch Series : Day 3/10
Every laugh and pause in this video was generated by our AI model.
This is Timbre 3, our new text-to-speech model, live today on the Gnani API.
Timbre 3 introduces expressions in your speech in the form of audio tags. You write a short cue where the delivery should change, and the voice performs it. There are 12 of them, covering laughter, breath, vocal style and pacing, across 11 Indian languages and many voices.
We built it for the places voice already does real work:
→ Voice agents and IVR that soften when they apologize and slow down when they read out an amount
→ The voice layer inside apps, from learning lessons to game characters
→ Creator productions, where the acting now lives in the script
Under the tags, Timbre 3 reads numbers and dates the way people say them, with natural prosody across every voice.
What would you make it say first?
#VoiceAI #TextToSpeech #Timbre #TTS
https://x.com/GnaniAi/status/2110288437580574918
Sarvam AIModel providermodel release2026-10-13
Sarvam ships Bulbul V4 Flash, a distilled Indian-language TTS model priced at ₹20 per 10,000 characters
The production cut of the Bulbul V4 line previewed in July (2026-07-30-sarvam-bulbul-v4-preview, W31) comes in a third below Bulbul V3 on price for Indian voice-agent traffic, a day after Raise put Cartesia Sonic on Dhan KYC calls (2026-10-12-raise-cartesia-dhan-voice); the speed claims are vendor-run with no external bench.
Details3 sources · 1 X post
Bulbul V4 Flash is a smaller real-time text-to-speech model distilled from Sarvam's V4 research, launching with 83 voices across 11 Indian languages on the existing Bulbul API.
Bulbul V4 Flash
KugelAudio Kugel 3.5 cuts TTS errors on short replies, numbers, and spelled codes at 25% below Kugel 3's price
Phone agents trip on digits, one-word replies, and booking codes more often than on prosody, and eight of the nine KugelBench tests score exactly those. Sarvam's same-day Bulbul V4 Flash (2026-10-13-sarvam-bulbul-v4-flash) sells normalisation of PNR and PAN numbers for the same reason (analytical). All figures are KugelAudio's own.
Details3 sources · 1 X post
KugelAudio's new default text-to-speech model, built on Kugel 3 for voice agents on phone calls, follows IPA dictionaries, timed pauses, and spell tags more closely and costs €0.0525 a minute through the same API.
Kugel 3.5
GradiumModel providerpartnership2026-10-13
Gradium TTS joins the Baseten Model Library as a one-click dedicated deployment
After LiveKit Inference (2026-09-09-gradium-livekit-inference, W37), Gradium's TTS now also runs as a dedicated deployment in a customer's own Baseten account, close to the self-hosted route PolyAI took to SageMaker a week earlier (2026-10-06-polyai-dialog-rsn-1-aws-sagemaker, analytical).
Details4 sources · 2 X posts
Teams running inference on Baseten can deploy Gradium's proprietary streaming TTS on dedicated H100 hardware and call it over the same WebSocket protocol as Gradium's hosted API.
Gradium TTS on Baseten · vendor Baseten
Our Text-to-Speech model is on the Baseten Model Library.
This means teams already running inference on Baseten can deploy Gradium TTS with one click alongside the models they already serve, with sub-100ms TTFA.
Point your agent at: https://docs.gradium.ai/guides/self-hosted/baseten-tts
https://x.com/GradiumAI/status/2110064061815689279
Gradium TTS by @GradiumAI is now in the Baseten Model Library, pairing ultra-low (50ms) TTFA with natural voices.
Use it with any transcription model and LLM to run your full voice agent pipeline on Baseten.
Try it here: https://www.baseten.co/library/gradium-tts/
https://x.com/baseten/status/2110058183293079597
Fish Audio releases Drama 3 with open weights, audio inpainting, and long-form voice consistency
Three weeks after the preview (2026-09-23-fish-audio-drama-3-preview, W39), the general release adds segment inpainting and downloadable weights under a research licence, and is pitched at audiobooks, dubbing, and games rather than live agents; its 80% launch discount lands the same day as Sarvam's cheaper Bulbul V4 Flash (2026-10-13-sarvam-bulbul-v4-flash), which targets a different market (analytical).
Details7 sources · 2 X posts
Drama 3, Fish Audio's directable text-to-speech model, is now generally available on fish.audio and its API, with weights on Hugging Face, taking plain-language performance notes and regenerating only a selected stretch of a take.
Drama 3
BolnaPlatformAPI2026-10-12
Bolna lets voice agents speak through any OpenAI-compatible TTS server
Teams running their own Indian-language voices can plug them in without waiting for a native integration, at the cost of owning TTS latency and uptime; Bolna already listed 12 native synthesizers on 14 September.
Also Connect
Details2 sources
Bolna customers can point an agent at their own text-to-speech endpoint, such as a self-hosted open-weight model or vLLM-Omni, and Bolna does not charge for that speech.
Bolna custom TTS
Showavatars, video7
TavusPlatformlaunch2026-10-16
Tavus opens a pilot of Kiosk mode, which puts a PAL on a walk-up kiosk
Square Kiosk added voice ordering at quick-service counters two days earlier (2026-10-14-square-kiosk-voice-ordering), and Tavus brings a face to the same walk-up setting; so far the only evidence is one demo at its own office.
Also Use
Details3 sources · 1 X post
Tavus is taking a limited number of companies into a pilot that runs its conversational video agents on walk-up kiosks, shown checking in a guest at its own front desk.
Tavus Kiosk mode
LemonSlice launches a Cartoon Model for real-time avatars of mascots, anime characters, and other non-human faces
Two weeks after CWM-1 Lite (2026-10-02-lemonslice-cwm-1-lite), LemonSlice now splits its avatar line by art style rather than price, aimed at games, children's apps, and brand mascots; the quality claims rest on the vendor's own demos.
Details2 sources · 1 X post
LemonSlice released a real-time avatar model tuned for cartoon characters, built from a single reference image and available on all subscription plans.
LemonSlice Cartoon Model
Introducing the LemonSlice Cartoon Model.
- Designed to bring cartoon characters to life with maximum animation
- Supports everything from 3D mascots to anime
- Highly expressive interactive avatars from a single image
Available today on all subscription plans! https://t.co/T1uF4SHnUd
https://x.com/LemonSliceAI/status/2111125116394332381
TavusPlatformlaunch2026-10-15
Tavus adds a ChatGPT plugin, Face Studio, and GPT-Live face rendering through both Pipecat and LiveKit
Avatar rendering is moving onto native speech-to-speech output, after Telnyx put preset faces on its AI Assistants on 13 October (2026-10-13-telnyx-ai-assistant-avatars-beta); the Tavus docs call the model GPT Live 1.5, while the newest OpenAI release in the corpus is GPT-Live-1.1 (2026-10-15-openai-gpt-live-1-1).
Also Orchestrate
Details4 sources
Tavus's changelog lets ChatGPT users build PALs through its MCP server, adds Face Studio for custom Phoenix-4.5 faces, and documents putting a Tavus face on OpenAI's GPT-Live speech-to-speech model.
Tavus ChatGPT plugin, Face Studio, and GPT-Live integration · vendor OpenAI
Kwindla Kramer shows a Halloween virtual pet running a small embedded Pipecat build
A character device running Pipecat on Deepgram speech is a preview of what the team plans to show at Deepgram Speak on 29 October; with the LLM undisclosed, the stack rests on Deepgram's one-line description.
Also Orchestrate
Details2 sources · 2 X posts
A cat character in a hardware case talks with Kwindla through what he calls a mini embedded Pipecat; Deepgram says it runs on Deepgram Flux STT and TTS with an unnamed LLM.
Halloween virtual pet (embedded Pipecat demo) · vendor Deepgram
I'll be at Deepgram Speak on October 29th with 500 or so new and old friends who build voice AI agents and applications. As mini embedded Pipecat and I were just discussing, I'm really excited about Deepgram Speak.
We're working on some very fun Pipecat stuff for the conference. If you're interested in models for realtime AI, building production voice agents, or what the future holds for voice and AI-native software, please join us in San Francisco.
https://x.com/kwindla/status/2110525021815001412
Check out @kwindla's Halloween virtual pet powered by Deepgram Flux STT + TTS and a super-secret LLM. 👇
See what else he's cooking up at Deepgram Speak, 10/29 in SF.
https://x.com/DeepgramAI/status/2110759855493652593
MarubunPlatformlaunch2026-10-15net-new
Kebbi Talk gives the Kebbi robot interruptible multilingual voice dialogue
Barge-in arrives on a robot Marubun has sold exclusively to Japanese care and medical customers since a March 2024 tie-up with NUWA Robotics, so existing Kebbi units gain interruptible dialogue; transcription, summaries, and alerts on worrying signs are still only planned.
Also Orchestrate
Details3 sources
Kebbi Talk adds natural voice dialogue to Marubun’s Kebbi communication robot, with adjustable role, voice, speed, and reference documents.
Kebbi Talk · vendor FriendlyTech
Gupshup launches Avatar Agent for self-serve video AI agents
Gupshup now sells chat, voice, and video agents from one self-serve line within six weeks (voice on 3 September); with no avatar, speech, or model vendor named, the rendering stack and per-minute cost are unknown.
Details3 sources
Businesses can build AI agents that talk to customers face to face through a digital avatar, answering questions, presenting products or webinars, and handing off to staff, with a free Gupshup-branded tier and pay-as-you-go pricing.
Avatar Agent
TelnyxPlatformlaunch2026-10-13
Telnyx gives its Voice AI Assistants six preset video avatars in beta, rendered on its own network
Telnyx now ships the face with the voice agent instead of leaving it to a separate avatar vendor such as Anam or Tavus, at $0.08 a minute from 1 November; the catch is that avatar calls run only on Telnyx's own TTS voices. Gupshup's Avatar Agent (2026-10-14-gupshup-avatar-agent) is a same-week comparison, not a shared integration.
Also Speak
Details1 source
Assistants can opt into video per conversation with one of six preset faces and three backgrounds, drawn by a Telnyx-hosted model that sits next to the assistant's voice, transcription, and tools.
Avatars for Telnyx Voice AI Assistants (beta)
ThinkLLMs, speech-to-speech7
Pipecat 1.13.0 adds Azure Voice Live speech-to-speech service
This brings the Azure Voice Live route alongside the earlier Microsoft Foundry voice agents (2026-09-24-microsoft-foundry-voice-agents); the Gemini (2026-09-15-google-gemini-3-8-live) and Nova (2026-10-05-amazon-nova-2-5-sonic) defaults change what unpinned bots run.
Details3 sources
AzureVoiceLiveLLMService supports function calls, caller transcription, semantic VAD, and Azure speech voices independently of Azure OpenAI Realtime.
pipecat-ai 1.13.0
Pipecat releases PhoneLLM Alpha 2 in Mini and Max sizes, with one-sentence reasoning on Max
Pipecat's own model now comes in a small and a large size and runs inside Flows, so teams can trade latency for multi-step accuracy without leaving the framework. OpenAI's removal of GPT-4o and GPT-4.1 snapshots in January gives Flows users still on those models a reason to test it now; the GPT-4.1 comparison is Daily's own. Continues Alpha 1 (2026-08-27-daily-pipecat-phonellm-alpha-1, W35).
Also Orchestrate
Details11 sources · 2 X posts
Two open-weight language models for voice-agent pipelines: a 30B Mini without reasoning for the lowest latency and cost, and a 120B Max that writes one sentence of reasoning before it answers, both trained on synthetic Pipecat Flows sessions.
PhoneLLM Alpha 2 (Mini and Max) · vendor NVIDIA
OpenAIModel providermodel release2026-10-15
OpenAI rolls out GPT-Live-1.1 to ChatGPT Voice with new voices and better accents
Five weeks after GPT-Live-1 reached the API (2026-09-10-openai-gpt-live-1), OpenAI is iterating the consumer model first; developers on the API, and on LiveKit's GPT-Live path (2026-10-15-livekit-agents-1-8-8-think), wait for 1.1 with no date, while native-language voices narrow the gap with regional TTS vendors in Hindi, Japanese, and Korean.
Also Speak
Details5 sources · 2 X posts
GPT-Live-1.1 replaces the speech-to-speech model behind Voice in ChatGPT, Codex, and dots for all users, adding native voices in French, German, Hindi, Japanese, Korean, and Mexican Spanish, more natural accents, and fewer filler phrases while it searches; API access is promised but undated.
GPT-Live-1.1
New voices. Better accents. And less "checking..."
GPT-Live-1.1 brings these updates to Voice in ChatGPT, Codex, and dots, with new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese.
Rolling out today to all users. Coming soon to the API.
https://x.com/OpenAI/status/2110797830361436604
GPT-Live-1.1 offers much better multilingual support (including several new voices), more natural delegation, and improved steerability. Rolling out now to ChatGPT Voice, Desktop, and Dots.
https://x.com/juberti/status/2110811122672795951
LiveKit Agents 1.8.8 runs GPT-Live through LiveKit Inference and swaps in Grok Voice 3.0
GPT-Live-1 (2026-09-10-openai-gpt-live-1) first reached LiveKit Agents through the 1.8.1 DuplexModel (2026-09-10-livekit-agents-1-8-1-duplex), which needed a separate OpenAI key; running it on LiveKit credentials puts duplex speech-to-speech under the same Inference account as the cascaded models. Grok Voice 3.0 (2026-10-14-xai-grok-voice-3-0) replaced Think Fast 2.0 in the model lists about a day after xAI's launch.
Details3 sources
livekit-agents 1.8.8 lets agents call OpenAI's GPT-Live 1 with LiveKit credentials through `inference.DuplexModel`, adds Gemini Live model strings to Inference, and replaces Grok Voice Think Fast 2.0 with Grok Voice 3.0 in the Inference and xAI plugin model lists.
livekit-agents 1.8.8
xAIModel providermodel release2026-10-14
Grok Voice 3.0 tops the Artificial Analysis speech-to-speech index at an unchanged $0.08 a minute
GPT-Live-1 still leads τ-Voice on xAI's own chart (74.5% against 71.6%), so the advantage is latency, duplex handling, and a $0.025 cost per task; teams pinned to grok-voice-latest got the new model without opting in, which matters for anyone running regression tests on a production agent.
Details2 sources · 1 X post
xAI's new speech-to-speech model now sits behind grok-voice-latest in the Voice Agent API for phone and app voice agents, at the same price as Grok Voice 2.0.
Grok Voice 3.0
Quansys AIModel providermodel release2026-10-13net-new
Quansys AI launches Atlas-v1, a half-duplex speech-to-speech model for on-premises voice agents
At ₹0.95 a minute for the full stack, Quansys prices well below hosted cascades, but a half-duplex model takes turns rather than listening while it speaks as GPT-Live-1 does (2026-09-10-openai-gpt-live-1); the Big Bench Audio ranking and sub-500 ms p99 remain vendor claims until results are published.
Also HearSpeak
Details2 sources
Atlas-v1 is a half-duplex speech-to-speech model that Quansys AI packages with its own speech and orchestration components for private deployment.
Quansys Atlas-v1
MOSIModel providerfunding2026-10-12
MOSI raises over RMB 1 billion for an all-modal real-time model
MOSS Live is still coming, while Pipecat’s alpha LLMWithBackend (2026-10-18-pipecat-1-13-0-orchestrate) ships a similar split between a frontend interaction model and a backend agent in the same week; this is an analytical comparison, not an integration.
Details3 sources
Huakong Fund led the Shanghai company’s Pre-A round to fund foundation-model R&D and global commercialisation.
MOSS Live
Orchestrateframeworks, evals24
Pipecat Context Hub 0.9.0 adds a repos command and a curated vendor repo catalogue for coding agents
A coding agent asked about a Tavus or AWS service now gets told when that vendor's code is not indexed instead of a confident answer without it; the previous hub release with new capability was 0.8.0 on 11 September (2026-09-11-pipecat-context-hub-0-8-0, W37), and the 15 October plugin (2026-10-15-pipecat-context-hub-plugin-0-2-3) calls this CLI, so existing indexes need `refresh --force` before the hints apply.
Details2 sources
Developers who ground coding agents in Pipecat through the hub can now add a vendor's repositories from a curated catalogue with one command, and searches say when a named vendor's code is missing from the index.
Pipecat Context Hub 0.9.0
Pipecat 1.13.0 adds frontend-to-backend LLM delegation
OpenAI Decisions (2026-10-15-openai-gpt-live-native-decisions) now has a Pipecat classifier path; LLMWithBackend remains alpha, so its split-worker design is not a general-availability claim.
Details3 sources
The alpha LLMWithBackend pairs a conversational frontend with a backend worker for deeper reasoning and tool orchestration.
pipecat-ai 1.13.0
Gnani launches Agentic Evals to score every voice agent before it takes a live call
Text-only simulation covers the logic of a call but not the audio, by Gnani's own account, and buyers cannot run it themselves yet; it follows AI Insights (2026-10-17-gnani-ai-insights) a day earlier, which reads the same agents' calls after they happen, and Talkative's test runs and graded calls (2026-10-15-talkative-voice-ai-editor) are the closest self-serve comparison this week (analytical).
Details2 sources · 1 X post
Automated pre-launch testing for the voice agents Gnani builds for enterprise customers: generated and custom test cases, simulated callers, and an LLM judge that blocks any agent scoring under 80 out of 100.
Agentic Evals
FlashAI Squads links specialised voice assistants into one team on a phone number
FlashLabs bases its claim that Squads is the first mid-call AI-to-AI handoff among Japanese voice-agent services on its own survey as of 18 October 2026, and concedes other vendors already sell agent platforms; the feature extends FlashAI 2.0 (2026-02-18-flashlabs-flashai-2-0).
Details2 sources
FlashAI Squads lets several specialised assistants transfer a live call between them while carrying its context forward.
FlashAI Squads
Gnani launches AI Insights to show why its voice agents' calls convert or stall
Topic shares come from a curated sample of answered calls, so small campaigns may show noisy splits; this is the second Gnani launch-series day filed this week after Timbre v3.0 (2026-10-14-gnani-timbre-v3), with Evon v3.3, Plexus early access, and the in-country deployments left unfiled.
Details2 sources · 1 X post
Enterprises running Gnani voice-agent campaigns get a funnel dashboard plus a view that sorts a sample of answered calls into Topics and Sub Topics and scores customer sentiment as positive, neutral, or negative.
AI Insights
Launch Series : Day 6/10
Introducing AI Insights, built to help enterprises understand their customers and the context behind what they ask.
#VoiceAI #AIInsights #EnterpriseAI #CustomerExperience https://t.co/701S1seRfx
https://x.com/GnaniAi/status/2111332629597294772
Gnani runs its voice-agent stack inside four sovereign regions
Gnani claims the whole runtime, not only storage, stays in-country in four regions, ahead of the Agent API it opened to partners three days later (2026-10-19-gnani-agent-api).
Also Connect
Details1 source
Gnani says its fully co-located AI stack is live in the US, Singapore, Japan, and the UAE, with production voice conversations running.
In-region agentic AI stack
Talkative moves Voice AI agents to a new editor with versioning, test runs, and graded call evaluations
Talkative switches off the old editor on 9 November 2026, so versioning and test runs become the default path for every Talkative voice agent; LiveKit Cloud's Agent Simulations reached GA a week earlier (2026-10-08-livekit-agent-simulations-ga, W41) with a similar scenario-run idea and no integration between the two.
Details2 sources
UK contact-centre software maker Talkative replaced its Voice AI configuration screen with one that versions agent changes and runs simulated test calls and reviews of real calls in the same place.
Voice AI Editor
Sarvam Voice Agents MCP lets coding assistants build, deploy, and test phone agents
Sarvam follows NiCE Cognigy (2026-08-26-cognigy-agentic-building, W35) and Bolna (2026-07-19-bolna-mcp-server-cli, W29) in handing agent building to coding assistants, and its tools reach live telephony: `place_test_call` dials a real number and resuming a campaign starts calling a cohort, which Sarvam's own docs tell users to approve call by call.
Details2 sources · 1 X post
A hosted MCP server at mcp.sarvam.ai/voice-agents gives Claude, Cursor, VS Code, and other MCP clients tools to configure Sarvam voice agents, connect telephony, run outbound campaigns and evals, and read call analytics.
Sarvam Voice Agents MCP
Introducing Sarvam Voice Agents MCP.
Describe the voice agent you want, and your coding tool builds it. The MCP connects your coding tool to Sarvam, so you can create agents, run campaigns, and analyse calls from the environment you already code in.
Watch Sarvam Voice Agents https://t.co/rFI5rOnikW
https://x.com/SarvamAI/status/2110703748083490882
Pipecat Context Hub plugin 0.2.3 packages Setup, Explore, Build, and Deploy workflows for coding-agent hosts
The plugin asks the user to install the Pipecat CLI (the CLI path from hub 0.2.0, 2026-06-13-pipecat-context-hub-0-2-0, W24) rather than relying on MCP, so a host without the CLI has no retrieval route until the user installs it.
Details2 sources
The first public Pipecat Context Hub plugin is a skills-only package of four workflows that call the hub's CLI for grounded Pipecat documentation, API definitions, and examples.
Pipecat Context Hub plugin 0.2.3
OpenAIModel providerAPI2026-10-15
OpenAI wires its Decisions API into GPT-Live so voice agents can act while the caller is still speaking
Routing and classification that voice agents usually run from their own backend after the caller stops can now run inside the Live session mid-utterance, though Decisions is still a beta tied to gpt-6-luna. On the framework side, LiveKit Agents 1.8.8 added a background decision-model plugin the same day (2026-10-15-livekit-agents-1-8-8-orchestrate); that is a comparison, not an integration.
Also Think
Details3 sources
GPT-Live sessions can now call the Decisions API natively, through delegation choices that Live hands off or through triggers on input and output transcript deltas, so apps get typed answers mid-turn without a round trip through their own server.
GPT-Live native Decisions integration
NineNineSix opens TYGER, a turn-taking model for full-duplex voice agents, with the tyger-s2s agent
Turn-taking ships as one open model that hears the agent's own voice, beside separate endpoint models such as Pipecat Smart Turn (2026-01-07-daily-pipecat-smart-turn-v3-2) and the LiveKit turn detector (2026-06-17-livekit-turn-detector-v1); its numbers are offline and self-reported, and it has not yet appeared on Cekura's Turn Bench (2026-10-13-cekura-turn-bench), published two days earlier.
Also Hear
Details4 sources · 2 X posts
Open weights for a speech decision model that transcribes the caller and, every 80 ms, chooses whether a voice agent should answer, keep listening, stop for a barge-in, or say "mm-hm", plus an Apache 2.0 browser agent built on it.
TYGER and tyger-s2s · vendor NVIDIA
LiveKit Agents 1.8.8 adds a DecisionModel for background choices, starting with TypeSafe Jev
Jev now has a hook in both Pipecat and LiveKit: Pipecat 1.12.0 added Jev evals on 25 September (2026-09-25-pipecat-1-12-0-classifiers-evals), and OpenCouncil (2026-10-14-opencouncil-voice-rooms) uses Jev for turn-taking. LiveKit adds it as a general `DecisionModel` slot, so other decision providers can use the same interface.
Details3 sources
livekit-agents 1.8.8 introduces `DecisionModel`, which runs probability, choice, and score evaluations in the background without delaying replies, with a first plugin for TypeSafe Jev (`jev-1.13.0`) and OpenRouter's Decisions API.
livekit-agents 1.8.8
LINE WORKS Catch Call answers business calls and sends recorded requests to staff
The 1,800-call annual tier costs ¥8,900 monthly after the free first year; LINE WORKS says its AiCall speech recognition handles 30 million calls a year.
Also Hear
Details3 sources · 1 X post
Catch Call takes the caller’s purpose, name, and callback details, then sends text and audio to a LINE WORKS room or email.
LINE WORKS Catch Call
AscendPlatformlaunch2026-10-15net-new
AgenFlow drafts and tests AI phone workflows and connects them to business systems
MCP support lets Claude and other assistants edit call flows and read call results, and plans start at ¥0 a month; live phone lines still need Ascend's separate AI Operator contract.
Also Connect
Details3 sources
AgenFlow handles inbound and outbound calls, completes after-call records, and can register information in CRM or booking systems.
AgenFlow
OpenCouncil puts several AI agents and a team in one live voice room, with Jev picking who speaks next
Several agents in one voice room need an arbiter for who talks next, and OpenCouncil buys that from Typesafe's Jev at about 370 ms, the model Pipecat demos used for character handover (2026-09-30-pipecat-multi-character-jev-handover-demo, W40). That is a shared vendor only; no Pipecat integration is named.
Also Hear
Details2 sources
OpenCouncil runs browser, Google Meet, and Discord voice rooms where agents on GPT, Gemini, Grok, or a team's own endpoint argue a question with the people present while a Scribe agent records decisions and next steps.
OpenCouncil · vendor Typesafe (Jev), LiveKit, and ElevenLabs
Inworld opens its open-model inference service to all developers, sized for voice-app first tokens
A TTS vendor now sells the language-model hop of a cascaded voice agent too, with a 117 ms median first token on Gemma 4 26B by its own measure; all comparisons are Inworld-run against OpenRouter medians, and the closest earlier filing is LiveKit's Gemma 4 inference (2026-07-02-livekit-gemma-4-inference).
Also Think
Details1 source · 1 X post
Inworld Inference serves open language models such as Gemma 4, GLM, DeepSeek, and MiMo for consumer apps, with per-model serving builds and speed or price tiers per request.
Inworld Inference
Inworld is your new favorite inference provider.
We already process tens of trillions of tokens a month because we understand consumer AI different, and we've built our inference from the ground up for it.
Today, the inference top consumer AI apps run on is opening up.
https://x.com/inworld/status/2110409542693478816
Pipecat demo adds sarcasm tags with SpellSpeak Tone RC4
Multi-character Pipecat agents can react to mock praise and ironic understatement without an extra LLM tool call — Tone RC4's sarcasm head runs in the same ~10 ms CPU pass as the existing emotion and act tags, with maker-reported 86% precision when it fires on game dialogue.
Details3 sources · 1 X post
Jon Taylor shows SpellSpeak Tone RC4, a ~28M-parameter ONNX MiniLM line classifier that adds a calibrated sarcasm head while keeping rc3 emotion, social-act, and conversational-move tags for multi-character Pipecat agents.
SpellSpeak Tone RC4 classifier demo
"Oh great, another release." New in Tone RC4: sarcasm detection 💀
Confidence output matters a lot here, as not all sarcasm is created equal, and when it calls sarcasm, it’s right 86% of the time (... or in the spirit of this tweet, 86% of the time, it's right every time!)
RC4 also comes as 16-bit weights now too at just 56 MB (half the download) with the same tags and the same speed on CPU.
Weights on HF: https://huggingface.co/spellspeak/tone
Demo: https://github.com/spellspeak/models/tree/main/tone/examples/pipecat
https://x.com/JonPTaylor/status/2110130003346444581
CrestaPlatformlaunch2026-10-13
Cresta launches Deputy Agent to carry customer-service workflows past the end of the call
A day after Sentry-0.1 (2026-10-12-cresta-sentry-0-1), Cresta pitches the weeks of follow-up work behind a call rather than containment on it; Aqua Finance and Sunbit are quoted on expected benefits, with no production figures yet.
Details3 sources · 1 X post
Deputy Agent is a long-running agent inside Cresta's CX platform that starts back-office work from live human or AI conversations and keeps it moving across systems, teams, and follow-up calls until the case closes.
Deputy Agent
Introducing Cresta Deputy Agent: a long-running AI agent that automates customer experience workflows from start to finish 🚀
From refund requests, to transaction disputes, to insurance claim appeals, many of the most important customer experience matters take days or weeks of manual work across systems and teams to resolve.
Deputy is purpose-built to handle this work, no matter how long it takes to complete.
Get all the details in our blog: https://cresta.com/blog/meet-deputy-agent-a-long-running-ai-agent-built-for-customer-operations
https://x.com/cresta/status/2110008055781806425
CekuraPlatformlaunch2026-10-13
Cekura launches Turn Bench across 11 realtime voice models
Teams choosing a duplex model for live calls get a public turn-taking board that separates quiet-room leaders from noise-stable ones — Gemini 3.1 Flash Live's finished-turn replies fall from 97% to 55% in heavy noise on Cekura's scripts, while GPT Realtime 2.1 barely moves, and no model both keeps going through "mm-hmm" and yields on real barge-ins.
Also Think
Details2 sources · 1 X post
Cekura's Turn Bench plays the same scripted callers to 11 realtime voice models and times four moments where the model must talk or stay quiet, including pauses, turn ends, barge-ins, and "mm-hmm" backchannels.
Turn Bench
Today we're launching @cekuraAi Turn Bench. We played the same scripted conversations to 11 realtime voice models and timed four moments where the AI has to choose between talking and staying quiet.
Most models handle the basics. The typical one replies 99% of the time once you've finished and stops every time you cut in. Pauses are harder: when you stop mid-sentence to think, it waits only 73% of the time.
People say "mm-hmm" to mean "keep going", but 9 of the 11 models stop talking almost every time they hear it. GPT-Live 1, the only model that usually keeps talking, also talks over most real interruptions. None of the 11 handles both.
Background noise changes the order. Gemini 3.1 Flash Live leads in a quiet room at 88.4%, but in heavy noise it replies to 55% of finished turns, down from 97%. GPT Realtime 2.1 moves by at most 3 points across all six audio conditions.
Pauses and turn ends are scored with Sesame's TurnBench scorer, unchanged. 15 of the scripts are public under CC BY 4.0.
http://benchmarks.cekura.ai/speech-to-speech/turn-bench
https://x.com/tarush_agarwal_/status/2110035317709558145
The Boston Agents Company launches Resia, a call and chat agent platform with orchestration agents
Multi-touch outreach (text first, call after a set wait, stop when the goal is met) ships as its own agent type with spend and call caps, where most voice platforms leave that sequencing to the builder's code; Vocaleo (2026-09-07-vocaleo-launch, W37) is the consumer product already running on it.
Also Connect
Details3 sources · 2 X posts
Resia lets builders run call agents, text and web-chat agents, and orchestration agents that combine both to chase one objective, from a portal or a single API request.
Resia
Today, The Boston Agents Company is announcing Resia.ai - the first voice agent platform built for multi-agent orchestration.
Resia enables anyone to build orchestration agents that coordinate conversations and actions across voice, web chat, SMS, and connected systems to drive objectives to the finish line.
It’s the system we wish we’d had when we started building with communications agents.
We built it to power our own products. Now it can power yours too.
https://x.com/PeterDaggett/status/2109992686824329265
Vocaleo was built on Resia - start building your own agents today at Resia.ai
https://x.com/vocaleo_agent/status/2109995205667758563
Bland rebuilds its dashboard around a unified agent builder
Bland folds pathways, tools, and knowledge into one agent node and sets a 15 November cut-off for production pathways and personas, so existing builds have about a month to move.
Also Connect
Details1 source · 1 X post
Dashboard v2 combines calls, SMS, webchat, and API interactions in one agent built from a single node with pathways, tools, and knowledge.
Dashboard v2
VapiPlatformlaunch2026-10-12
Vapi adds traffic splitting and rollback for assistant versions
Version-aware logs make it possible to inspect canary calls by release; instant rollback limits exposure if a new assistant version fails.
Details1 source
Vapi can route a percentage of traffic across published assistant versions, supporting canary rollouts and instant rollback.
Assistant version traffic splitting
telliPlatformlaunch2026-10-12
telli adds Simulations to test voice agents before live calls
Built-in simulation puts pre-launch testing inside the agent builder, the same job as Gnani Agentic Evals (2026-10-18-gnani-agentic-evals) and Talkative Test Runs (2026-07-16-talkative-voice-ai-test-runs-beta) (comparison only); telli dates the entry to the week of 12 October 2026, not a single ship day.
Also Use
Details2 sources
The new page lets teams create reusable scenarios, select an agent, version, and conversation count, then inspect each run and conversation side by side.
Simulations
CrestaPlatformmodel release2026-10-12
Cresta launches Sentry-0.1 to flag personal AI agents in customer conversations
Cresta reads conversational behaviour rather than audio and gives an F1 of about 0.9 from preliminary evals, the first accuracy figure in the agent-caller detection thread: Sierra Fleming (2026-10-06-sierra-curie-fleming-models), ElevenLabs synthetic voice detection (2026-10-09-elevenlabs-synthetic-voice-detection), Liberate AI Intercept, and Decagon's Personal Agent Gateway published none.
Details2 sources · 1 X post
Cresta's real-time ensemble model, built into its CX platform, scores each conversation for behavioural signs that the customer is represented by a personal AI agent, so businesses can guide human reps, apply guardrails to agent-to-agent exchanges, and analyse why agents get in touch.
Cresta Sentry-0.1
We’re introducing Cresta Sentry-0.1, a model to detect personal agents in conversations.
As consumers delegate more of their to-do list to personal AI agents, Sentry enables businesses to guide human reps, apply guardrails to agent-to-agent interactions, and analyze why agents are calling.
Learn more: https://cresta.com/blog/introducing-cresta-sentry-0-1-recognize-personal-ai-agents-and-guide-every-response
https://x.com/cresta/status/2109770896701305235
Connecttelephony, channels9
Pipecat 1.13.0 adds RTVI file uploads and video source capture
The server-side RTVI 2.2.0 media path follows client transport media support (2026-10-16-pipecat-client-transports-media-support); applications must configure upload storage before offering the development upload route.
Also See
Details3 sources
RTVI 2.2.0 accepts client send-file messages for images and files, with resolver-backed LLM context and bot backchannels exposed to clients.
pipecat-ai 1.13.0
TelnyxPlatformlaunch2026-10-16
Telnyx opens Programmable Voice in Mumbai with local media anchoring for Indian voice agents
Indian voice-agent builders on Telnyx can keep numbers, media, speech models, and inference on one in-country network, removing the Singapore or Sydney hop that Telnyx names as the old latency cost; it gives no millisecond figures. Extends the Chennai anchorsite (2026-09-30-telnyx-chennai-sip-anchorsite, W40).
Details3 sources · 1 X post
Voice API and Call Control applications can now anchor call media in Mumbai, so builders serving Indian callers no longer route media through Singapore or Sydney.
Programmable Voice in Mumbai
India's voice AI stack is coming together.
With Programmable Voice now available in Mumbai, developers can bring together local telephony, Indian mobile numbers, speech models, and AI inference on the Telnyx platform.
Whether you're building voice AI agents for Indian consumers or scaling contact centre applications, the building blocks are now in place to run voice experiences for India on one network.
One platform. From the phone number to the AI.
https://x.com/telnyx/status/2111083260076019843
Pipecat client transports report mediaSupport so UIs can hide controls that cannot work
A voice UI can drop the camera or screen-share button on audio-only WebSocket and MoQ sessions instead of showing controls that fail; all five transports shipped within 16 to 17 October 2026 and need client-js 1.14.0.
Details6 sources
Daily, SmallWebRTC, WebSocket, LiveKit and MoQ client transports each implement a mediaSupport getter that PipecatClient combines with what the bot supports; it needs client-js 1.14.0 or later.
Pipecat client web transports (mediaSupport)
FlukePlatformlaunch2026-10-16net-new
Fluke bundles AI voice agents with a phone line for Brazilian businesses, built without code, by API, or over MCP
Fluke sells the number and the agent on one bill and lets Claude or ChatGPT create both over MCP, the fourth voice MCP server in W42 after Gladia (2026-10-13-gladia-mcp-server), Viirtue (2026-10-14-viirtue-mcp), and Sarvam (2026-10-15-sarvam-voice-agents-mcp); it names no speech model, and only two Brazilian firms were testing in September.
Also Orchestrate
Details5 sources
Brazilian digital mobile operator Fluke opened voice agents that answer and place calls from a Fluke number and hand over to staff, on plans from R$199 a month in its Fluke Empresas portal.
Agentes de voz (Fluke Empresas)
ParloaPlatformlaunch2026-10-15
Parloa introduces Personal Agent Middleware for personal AI agents
Parloa follows Vapi Personal Agent Calling (2026-09-30-vapi-personal-agent-calling, 30 September) and Decagon Personal Agent Gateway (2026-10-01-decagon-personal-agent-gateway, 1 October) in building the business-side entry point for Muse-class personal agents (comparison, not a documented link); access is limited to design partners, with no general availability date.
Also Orchestrate
Details2 sources · 1 X post
The limited design partnership connects a customer's personal agent to a brand's systems while checking identity, permissions, risk, and approvals.
Personal Agent Middleware
What happens when your customer’s AI agent contacts your business?
Last week, we introduced Parloa’s Personal Agent Middleware. Watch the demo to get a sneak peek at how it works.
Personal agents connect. Customers approve. Businesses stay in control. https://t.co/x39ownyugD
https://x.com/parloa_ai/status/2112528554911260986
BolnaPlatformlaunch2026-10-15
Bolna SIP trunks connect voice agents to customer carriers
The shared channel cap and carrier credentials shift capacity and routing control to the customer; this extends the earlier SIP transfer (2026-08-25-bolna-sip-trunk-transfer) alongside Bolna’s bring-your-own TTS (2026-10-12-bolna-byo-tts).
Details2 sources
Customers can route inbound and outbound Bolna agent calls through Plivo, Telnyx, Twilio, Exotel, Vobiz, Vonage, or another standards-based SIP carrier.
Bolna SIP trunks
Bland builds a branded iMessage business agent from a website link in about 30 seconds
iMessage goes from an Enterprise add-on (7 May) to a 30-second self-serve agent two days after Bland rebuilt its dashboard around one agent across phone, messaging, web, and API (2026-10-13-bland-dashboard-v2), so the agent that answers a company's calls can also hold its branded thread on customers' iPhones; the channel is text-first, with voice limited to memos.
Also Orchestrate
Details3 sources · 2 X posts
Business Agents gives a company its own iMessage contact, with its name and logo, that Bland generates from the company's site and wires to its APIs for customer conversations and in-chat payments.
Business Agents (iMessage)
Viirtue MCP lets Claude, ChatGPT, and other assistants query a reseller's phone system and AI Voice Agents
Voice-agent call logs and resolution rates become readable from a general-purpose assistant without a vendor portal; Viirtue is the third voice MCP server this week after Gladia (2026-10-13-gladia-mcp-server) and Sarvam (2026-10-15-sarvam-voice-agents-mcp), but it is read-mostly and aimed at office managers rather than agent builders.
Details3 sources · 1 X post
White-label VoIP platform Viirtue opened a remote MCP server to its MSP and telecom-reseller partners, so their customers can ask an AI assistant about calls, queues, voicemail, fax, and voice-agent results under the partner's brand.
Viirtue MCP
Missed the big reveal? Catch the replay. 🎬
Yesterday we introduced Viirtue MCP, a new way to bring your phone system inside the AI you already use.
Viirtue MCP connects Claude, ChatGPT, and other AI assistants to your cloud phone system, call center, AI Voice Agents, and fax. https://t.co/zqvFS5GFwG
https://x.com/Viirtue1/status/2110747820781126066
Cloudonix adds a contact-centre edition to OPBX, its open-source phone system for AI voice agents
The release extends the May OPBX launch (2026-05-26-cloudonix-opbx-launch, W22) to mixed human and AI contact centres but lists no capability beyond that framing, so it reads as positioning until feature docs follow.
Details3 sources
OPBX Contact Center puts human agents and AI voice agents on one open-source, multi-tenant phone system built on the Cloudonix CPaaS, offered hosted or self-hosted.
OPBX Contact Center
Useagents in market32
SocialRoots.ai adds an AI Agentic Caller to GridSocial for Medicaid renewal outreach
The 1 January 2027 requirements give this outreach a fixed deadline, but SocialRoots.ai names no health centre customer or call volumes to establish deployment scale.
Also Connect
Details3 sources
FQHCs and community health organisations can use multilingual calls and texts to check renewal status where connected data is available, screen household changes, and connect members to enrolment support.
AI Agentic Caller (GridSocial)
walkai-talkai lets developers speak to Cursor or Claude Code through a handheld two-way radio and hear the agent's reply over the air
A radio channel has no speaker check, so anyone transmitting on it can reach the agent; the maker replied on 18 October that prompt injection over radio has probably not been studied yet. Point and Yap (2026-10-18-point-and-yap-screen-pointing-mcp) is the same week's other voice front end for coding agents, using a local MCP queue instead (analytical).
Also HearSpeak
Details3 sources · 2 X posts
A paid macOS menu-bar app bridges any VOX-capable walkie-talkie to coding agents, transcribing radio speech on the Mac and reading the agent's short summary back into the radio channel.
walkai-talkai
Emniyette görev yaparken telsiz kullanmayı severdim. Çünkü çok hızlı bir iletişim aracı. Bir hafta önce aklıma ilginç bir fikir düştü: Acaba yapay zekayla telsiz üzerinden konuşabilir miydik?
Bunun için güzel bir isim seçtim ve bir uygulama geliştirdim: https://t.co/Y3JzWSvIk4 https://t.co/MnHuQ75GhM
https://x.com/sezeriltekin/status/2111475969550627173
we’re going live on product hunt! 🚀
i’d really appreciate your support for walkai-talkai ❤️
if you have a moment, leave an upvote, and share your thoughts. it would mean a lot to us!
and yes, there’s 50% off for those who want to give it a try 👀
🔗 https://t.co/Y3JzWSwg9C https://t.co/bf7Jn4DMPr
https://x.com/sezeriltekin/status/2112140504435691561
ParrotPlatformlaunch2026-10-17net-new
Parrot records Mac calls without a bot and surfaces answers from the user's own documents while the client is still talking
Nani? (2026-10-16-coretech-labs-nani-live-meeting-subtitles) and ParakeetAI 2.0 (2026-10-07-parakeetai-2-0, W41) also listen beside a call and prompt the user mid-conversation; Parrot is the only one of the three with open source code, and its on-device-only profile lets a user keep a sensitive call away from any cloud model (analytical). It is unrelated to Ringg's Parrot STT model (2026-06-05-ringg-parrot-stt, W23).
Also Hear
Details4 sources
The free, GPL-3.0 macOS app captures system audio and the microphone, transcribes and separates speakers on the device, and runs an assistant on Claude, a local Ollama model, or another provider of the user's choice that cites the source file for each answer and writes a report of promises and missed questions after the call.
Parrot 0.29.7
Kente Labs' Emochi turns spoken diary entries into a shape-shifting character that maps the day's feelings
Hume's relaunched Expression Measurement API (2026-10-06-hume-expression-measurement-api, W41) scores emotion from the voice signal, whereas Emochi's listings describe reading feelings from transcribed words, which puts it closer to speech-in capture apps such as Taim (2026-08-04-taim-voice-first-thought-capture, W32) than to affect models (analytical); the 25-day report gate means its pattern claims are untested at launch.
Also Hear
Details3 sources
The free iPhone journal transcribes voice notes, lets users edit the text, and shows every emotion it reads in an entry as the shape of a character called Mochi, with topic threads, a calendar of trends, and a personal report after about a month.
Emochi
Voip Unlimited launches AI Agent, a speech-to-speech phone agent that channel partners build, brand, and bill themselves
A second UK carrier now sells a white-label phone agent to resellers, after Voiceflex VoiceAI (2026-10-09-voiceflex-voiceai, W41); neither names the speech-to-speech model underneath, so latency and language claims cannot yet be checked.
Also Connect
Details4 sources
UK wholesale carrier Voip Unlimited opened AI Agent to resellers through its Partner Portal to answer missed, out-of-hours, peak, and routine calls for their business customers.
AI Agent
NTT Docomo Business trials its CONN voice assistant on ElevenAgents with older residents of snowbound Aizuwakamatsu
A Japanese carrier's enterprise arm now runs ElevenAgents in a regional health trial for older residents as well as in contact-centre pilots, though the trial has run since autumn 2025 and the story gives no call volumes. Banner Health's scheduling agent (2026-10-02-banner-health-elevenagents-scheduling, W40) is the nearest ElevenAgents healthcare deployment (analytical).
Also Speak
Details1 source · 1 X post
With Takeda Health Foundation and Aizuwakamatsu City, the enterprise arm of NTT Docomo runs a voice agent that gives older residents in Fukushima local information and everyday conversation.
CONN on ElevenAgents · vendor ElevenLabs
NTT Docomo Business is trialing a voice agent for older residents of Aizuwakamatsu, Fukushima, built on ElevenAgents.
Heavy snow can make it hard for older residents to leave home in winter. With Takeda Health Foundation and the cooperation of Aizuwakamatsu City, NTT Docomo Business deployed CONN, its AI assistant, to share local information and hold everyday conversations.
Read the full story: https://t.co/o6d3n8kTGh
https://x.com/ElevenLabs/status/2111098257225236768
NTAPlatformlaunch2026-10-16net-new
NTA builds AgenteX, an AI receptionist that reads extension status and transfers calls inside its hosted phone system
NTA is the third UK carrier to add a phone agent within a week, after Voiceflex VoiceAI (2026-10-09-voiceflex-voiceai, W41) and Voip Unlimited AI Agent (2026-10-16-voip-unlimited-ai-agent); it pitches extension status and call history read from inside the switch, but like the other two names no speech model or price.
Also Connect
Details2 sources
UK VoIP carrier NTA added AgenteX to its core telephony platform, so the AI receptionist answers inbound calls with access to live call activity, staff availability, and earlier conversations.
AgenteX
Coretech Labs' Nani? floats live translated subtitles over any Mac call and chimes when a question is aimed at you
Zendesk (2026-09-10-zendesk-realtime-voice-translation, W37) and Konecta (2026-09-24-konecta-v2v-voice-translation, W39) build translation into the contact centre, whereas Nani? runs on one listener's Mac with no bot in the call, so other participants may never know it is there; the FAQ leaves any duty to tell them to the user (analytical).
Also Hear
Details3 sources
The macOS app listens to the audio the Mac plays without joining the meeting, labels speakers by voice on the device, and suggests replies the user can say back, starting with Japanese and English among nine languages.
Nani?
Bank BTNEnterpriselaunch2026-10-16net-new
Bank BTN puts Jasnita's AI Talkbot on outbound payment-reminder calls
Collections reminders are the commonest outbound voice-agent job filed this month, alongside Insight7 (2026-10-15-insight7-voice-ai-agents) and PayNearMe (2026-10-01-paynearme-ai-servicing-collections-agent, W40); BTN's release has no call volumes or results, so only the adoption by an Indonesian state bank is on record.
Details3 sources
Jasnita Telekomindo supplies the Indonesian state housing bank with a voice bot that phones borrowers with payment reminders and due dates, leaving negotiation and special cases to staff.
AI Talkbot for desk collection · vendor Jasnita Telekomindo
Answering Service Care's AI assistant now places scheduled outbound calls for collections, confirmations, and reminders
A human answering service adding AI outbound lands the same week as Ada Proactive Outreach (2026-10-13-ada-proactive-outreach-voice) and Insight7 (2026-10-15-insight7-voice-ai-agents); the claim that AMD Premium gets past iPhone call screening is the vendor's own and unverified.
Also Connect
Details1 source
Accounts with AI answering can set an AI Call To Do once, on a recurring schedule, or from their own software, and the assistant phones the customer, holds the conversation, and reports how it went.
AI Call To Dos
SBI Shinsei Bank joins its AI phone line to the contact centre so callers can switch to an operator mid-call
Operator handoff arrives eight months after the Bright 60 launch, whose proof of concept scored 99% on about 3,000 FAQ calls; with every transaction still routed to staff, the AI line stays an information tier for a bank that takes about 300,000 phone enquiries a year.
Also Orchestrate
Details2 sources
Since 2 October the Japanese bank's natural-conversation voice AI line has answered a wider set of retail questions and passed callers to a human operator on request, with procedures and transactions left to staff.
SBI Shinsei Bank AI Window (SBI新生銀行AI窓口) · vendor Recho
Insight7 launches inbound and outbound voice agents for collections, reminders, and CRM updates
A call-analytics vendor now sells agents that place the calls as well as score them, and Nigerian Pidgin support is rare among the outbound collections products in this corpus.
Also Orchestrate
Details1 source
Call-analytics company Insight7 now sells voice agents that take inbound calls and run outbound campaigns for payment reminders, collections, application follow-ups, and scheduling, with English, Nigerian Pidgin, and other Nigerian languages.
Insight7 Voice AI
GeelyEnterprisepartnership2026-10-15net-new
Geely takes Cerence AI's multilingual speech from its car cockpits into service robots
An in-car voice supplier following a carmaker client into robots. This is a motor-show debut with no robot model, site, or date, unlike the dated Hyundai BAYON go-live with SoundHound (2026-10-12-hyundai-motor-india-bayon-soundhound) this week. Cerence's earlier Geely Auto cockpit deal is 2026-01-05-cerence-geely-auto-xui (W02).
Also Hear
Details2 sources
Cerence AI's speech recognition and conversational AI move from Geely's intelligent cockpits into service robotics, shown publicly for the first time at the Paris Motor Show with French, English, and Mandarin recognised at once.
Cerence AI speech for Geely service robots · vendor Cerence AI
AtticusPlatformlaunch2026-10-15net-new
Atticus turns a held shortcut into spoken commands that run apps, windows, files, and web errands on the Mac
Billing every ask to an existing Claude subscription leaves the maker with no inference bill, which is how a free open-source app can offer the hold-to-talk desktop agent that Incredible (2026-10-06-incredible-voice-computer-agent, W41) sells after a 7-day trial (analytical). The cost and speed figures are the maker's own, and the macOS 26 floor comes from Apple's speech model.
Also HearSee
Details4 sources
Atticus is a free, MIT-licensed Mac assistant: the user holds Control, Option, and Space, says the task, and it opens apps, arranges windows, reads the screen, drafts replies, or hands larger jobs to background agents, running on the Claude Pro or Max plan the user already pays for.
Atticus 1.2 for Mac · vendor Anthropic Claude (through Claude Code) and Apple on-device speech recognition
AloTechEnablerpartnership2026-10-15net-new
Turkish CCaaS AloTech names Parloa its strategic AI partner for contact-centre agents
Parloa has signed two channel partners in two days, TELUS Digital for implementation and AloTech for Turkish CCaaS customers, so it is growing through partners that already hold the contact-centre contracts rather than selling every enterprise directly.
Also Connect
Details1 source
AloTech, which calls itself Türkiye's largest cloud contact-centre platform, will recommend and deploy Parloa's AI agents to its enterprise customers, Parloa's route into the Turkish market.
Parloa agentic AI for AloTech customers · vendor Parloa
TELUS Digital becomes an implementation partner for Parloa's AI agents
Parloa's own 2026 ships are now filed under the `parloa` company, from the Series D (2026-01-15-parloa-series-d) to Navigator and Lens (2026-06-23-parloa-navigator); the TELUS release names no client, metric, or split between voice and chat volume.
Also Orchestrate
Details4 sources
Enterprises buy Parloa's agent platform directly and hire TELUS Digital for integration, change management, managed services, and optimisation, while TELUS Digital introduces Parloa to its clients in North America and Europe.
Parloa implementation partnership · vendor Parloa · delivered by TELUS Digital
SquarePlatformlaunch2026-10-14net-new
Square Kiosk takes quick-service restaurant orders by voice as well as touch
A point-of-sale vendor with its own hardware and payments now bundles voice ordering at $99 a month per device, which brings it into the restaurant market where the QSR specialists on the vertical roster (Presto, SoundHound, ConverseNow) sell; Square publishes no accuracy figures and names no speech vendor.
Details2 sources · 1 X post
Square's new self-service kiosk for quick-service restaurants adds a conversational voice layer that knows the seller's menu, answers questions, handles modifications and substitutions, and suggests add-ons.
Square Kiosk
Meet the new Square Kiosk, smarter self-service for a faster rush. Guests can browse by touch, ask menu questions by voice, and order on a 21.5-inch display built for busy restaurants. It’s self-service, reimagined.
https://x.com/Square/status/2110401688347570440
Screenwell's Luna runs first-round candidate screenings as a spoken video call and ranks applicants with quoted evidence
Per-screening pricing puts the AI interviewer against the recruiter's cost of a 30-minute phone screen. Toby (2026-10-13-toby-mock-interviews) is the candidate-side rehearsal of the same call, and Underheard (2026-10-07-underheard-ai-interviewer-ga, W41) aims a voice interviewer at customers (both analytical).
Also Show
Details2 sources
Recruiters send applicants a link from the job ad, or add Luna to a Google Meet or Teams call, and the lip-synced AI interviewer holds a short screening in the applicant's language and returns a scored, ranked shortlist.
Screenwell (Luna)
Confido Health raises a $12.5M Series A extension led by Z47 to widen CareOS beyond voice agents
A week after Vocca's $20M Series A (2026-10-06-vocca-series-a) for practice phone agents, Confido raises on the pitch that calls are only the first step. Its 95% handle rate and doubled collections are company or customer figures with no call volumes behind them.
Details3 sources · 1 X post
The New York company, which began with voice agents for healthcare, takes total Series A funding to $22.5M to turn CareOS into an operating layer that it says runs a practice from end to end, aimed at multi-site provider groups.
CareOS
We've raised $22.5M in total Series A funding for Confido Health, including a new $12.5M extension led by @z47_vc.
Medicine has added decades to our lives, yet patients still lose weeks to paperwork.
We're building AI agents to give healthcare teams more capacity to care. https://t.co/1iMJzV3G4U
https://x.com/mchetanreddy500/status/2110363317382369453
CMA CGM scales HappyRobot voice agents across sales, customer service, and finance
A carrier with more than 54 million customer requests a year publishes outbound sales results (about a third of 4,000 contacted customers moved to route talks) next to 46% inbound resolution, and it also holds equity in its vendor; Samsara (2026-10-13-samsara-agent-studio-ga) is the same week's other logistics voice-agent ship, with no link between them.
Details2 sources
The shipping and logistics group is taking HappyRobot voice agents from pilots to wider use, including Eva, which phones registered customers who have no sales rep, plus agents that answer cargo-release queries and chase unpaid invoices.
HappyRobot voice agents (Eva) · vendor HappyRobot
WSECU moves member calls to Talkdesk, where its Ava agent contains 80% of routine requests
A US credit union reporting 80% containment sits beside Raise's Cartesia KYC bot (2026-10-12-raise-cartesia-dhan-voice) as a same-week financial-services voice deployment; the containment figure is Talkdesk's and the story gives no period or baseline.
Details2 sources · 1 X post
Washington State Employees Credit Union replaced an on-premises phone system with Talkdesk's banking cloud, where voice-authenticated members can hand balance checks, transaction reviews, and other routine requests to its agentic AI agent Ava.
Ava (Talkdesk Customer Experience Automation) · vendor Talkdesk
By combining Talkdesk Customer Experience Automation (CXA), agentic AI, & a hybrid workforce, WSECU is making routine banking faster & easier while giving employees more time to focus on complex member needs.
Check out our new WSECU case study below:
https://www.talkdesk.com/customers/wsecu/
https://x.com/Talkdesk/status/2110022829311664525
WebFXEnterprisepartnership2026-10-13net-new
WebFX runs first-round recruiting screens through a Fireflies voice agent
Recruiting was one of three launch uses for Fireflies Voice Agents (2026-09-02-fireflies-voice-agents, W36), and WebFX's screens ran from January, so the agent was with customers months before that September filing; the case study's own time-saved figures disagree.
Details2 sources · 1 X post
Digital marketing agency WebFX offers job candidates an immediate screening call with a Fireflies Voice Agent that asks its own interview questions, then routes transcripts, summaries, and scorecards to recruiters.
Fireflies Voice Agent (screening interviews) · vendor Fireflies.ai
1 in 3 hires at WebFX came from a Fireflies Voice Agent screening.
Their recruiting team was stuck with a three-week wait for first-round screens.
Now candidates can choose to talk to a Voice Agent immediately. 277 screenings later, that's 83+ hours back for the team, and a lot fewer candidates lost to other offers.
The full WebFX Case study: https://fireflies.ai/blog/webfx-fireflies-voice-agents
https://x.com/firefliesai/status/2110008626026856950
TobyPlatformlaunch2026-10-13net-new
Toby runs spoken mock interviews for a specific job on a call styled like Meet, Teams, or Zoom
Per-minute pricing with no subscription ties the cost to live voice time. HireKey (2026-10-11-hirekey-voice-interview-coach, W41) bundles a voice interview coach inside a job-search suite, while Toby sells only the role-specific spoken rehearsal (analytical).
Also HearShow
Details3 sources
Paste a job description and Toby researches the company, maps the likely rounds, and interviews the candidate by voice as 'Toby Higgins', then returns a scored debrief with pace, filler words, and the transcript.
Toby
Samsara opens Agent Studio, with voice agents that call drivers, to all customers
Outbound voice reaches commercial drivers through the in-cab device rather than a handset, which suits US rules that allow only hands-free phone use while driving; the claim that voice passed chat within four weeks is a rare published channel split from a deployed agent builder, but Samsara gives no absolute call volume.
Also Orchestrate
Details3 sources · 1 X post
Samsara's no-code builder for chat and voice agents running on fleet and site data is now available to every Samsara customer, included in existing subscriptions.
Samsara Agent Studio
Samsara Agent Studio is now available to all customers.
Built for physical operations, it lets teams create AI agents that automate reporting, identify fuel savings, anticipate compliance risk, and support drivers through voice: https://www.samsara.com/blog/turning-operational-data-into-action-with-samsara-ai
https://x.com/Samsara/status/2110003993946603848
KakaoPlatformlaunch2026-10-13
Kakao unveils Kanana Space, a voice-first KakaoTalk assistant fronted by the Ryan character
Kakao pitches the assistant at roughly 50 million KakaoTalk users, but nothing reaches them before the December closed beta, and its full-duplex Kanana Live mode is still in development while Decagon Voice 3 (2026-10-01-decagon-voice-3-chord) already ships a duplex architecture; the voice stack traces to Kanana-o's instructable speech upgrade (2026-08-04-kakao-kanana-o-instruct-tts, W32).
Also Think
Details5 sources
Kanana Space opens from a two-finger swipe anywhere in KakaoTalk and takes spoken or typed requests to send messages, manage schedules, search the web, and change app settings, with pre-registration from 13 October and a closed beta in December.
Kanana Space
Josh Software offers banks a voice banking build they own for regional-language customers
No bank is named, so the SI stays the filed actor; handing source code to the bank differs from vendor-run Indian BFSI voice agents such as Credgenics Prix AI (2026-10-03-credgenics-prix-ai) and Desible.ai (2026-10-07-desible-seed-plus-funding), which sell an operated platform.
Also Orchestrate
Details2 sources
Pune engineering firm Josh Software launched a service that wires voice AI and intent recognition into banks' existing systems for Tier II, Tier III, and rural customers, handing over the source code so each bank keeps the IP.
Voice banking service
DaptaPlatformfunding2026-10-13
Dapta raises a $6M seed for its SMB voice and text sales agents
Spanish-first distribution through Dapta's own social content, about 4,000 sign-ups a week by company count, is the growth engine rather than enterprise sales; Resia (2026-10-13-boston-agents-resia) is a same-day platform that also mixes calls and texts for follow-up, with no link between them.
Also Orchestrate
Details4 sources
Chingona Ventures led the round in Dapta, whose Dapti agent answers inbound leads by phone, text, or WhatsApp in English or Spanish, qualifies them, and books appointments for small service businesses.
Dapti
ApplePlatformlaunch2026-10-13
Apple HomeView brings Siri AI to a shared home display run mostly by voice
HomeView takes Siri AI, in beta on iPhone since 14 September (2026-09-14-apple-siri-ai), into the shared-room role that Alexa+ holds on Echo devices (2026-09-16-amazon-alexa-plus-india). The assistant is still an English-only beta, and Apple gives no voice latency or recognition figures.
Details1 source
Apple's $299 7-inch tabletop or wall-mounted hub recognises household members by face to unlock their personal Siri context, controls the smart home hands-free, and continues conversations in the Siri app on other Apple devices.
HomeView
AdaPlatformlaunch2026-10-13
Ada Proactive Outreach lets its AI agent place outbound customer-service calls
Outbound calling now runs on the same Ada agent, knowledge, and Playbooks as inbound, over the customer's own trunk and caller ID. Vapi Campaigns (2026-08-31-vapi-vapi-campaigns) and Voiso (2026-09-23-voiso-ai-voice-agents) reached outbound earlier; Ada publishes no customer, price, or resolution figure yet.
Also Connect
Details1 source · 1 X post
The same Ada agent that answers inbound calls can now phone customers first, either through Ada's API over the business's outbound SIP trunk or by taking calls a contact-centre platform has already connected, then run a Playbook such as rebooking a cancelled flight.
Proactive Outreach (voice)
Your AI customer service agent can now make the call in the moments that drive loyalty. With Ada Proactive Outreach, your AI agent can reach customers before they have to reach you, and resolve directly on the call. Watch the full demo: https://t.co/QfLAQpwyKY https://t.co/1msWlMbsX2
https://x.com/ada_cx/status/2110398209860182022
Raise Financial Services puts Cartesia Sonic on Dhan voice bots for multilingual KYC
A major Indian brokerage group has Cartesia TTS on production Dhan voice KYC with three-quarters of those calls in Indic languages — Cartesia's customer story, not a new Sonic model release.
Also Speak
Details1 source
Cartesia said Raise is running Sonic on a share of Dhan voice-bot calls, including conversational KYC, with more than 75% of those interactions in languages other than English.
Dhan voice bots (Cartesia Sonic) · vendor Cartesia
KudosPlatformlaunch2026-10-12net-new
Kudos 3.0 rebuilds its wallet around an AI that phones providers to cut bills and cancel subscriptions
ClawCall (2026-10-05-clawcall-chatgpt-plugin) and Tab (2026-10-06-tab-personal-agent-calling) sell outbound calling as an add-on to personal agents; Kudos embeds the same pattern inside a card-wallet app for bill and subscription chores, with vendor-claimed ~$216/yr per negotiation and a prior Premium bill-negotiation SKU already on Product Hunt earlier in 2026.
Also Connect
Details3 sources
The consumer finance app's Product Hunt 3.0 launch centres on an AI agent that dials mobile, ISP, TV and security providers, negotiates rates within the existing plan, cancels forgotten subscriptions, and returns call recordings to the member.
Kudos 3.0
Hyundai Motor India puts a SoundHound and Perplexity generative voice assistant in the BAYON SUV
A second Korean carmaker in India has picked SoundHound for a generative in-car assistant within a fortnight of Kia's Sorento (2026-09-30-kia-india-soundhound-sorento, W40), and this one names Perplexity for answers, so SoundHound supplies the speech layer while a third party supplies the knowledge.
Also ThinkHear
Details4 sources · 2 X posts
Hyundai Motor India's first generative AI voice assistant, woken with 'Hey Hyundai', ships in the India-spec BAYON with voice AI from SoundHound and answers from Perplexity, alongside more than 500 embedded offline commands in six languages.
Hyundai BAYON Generative AI Voice Assistant · vendor SoundHound AI (voice) with Perplexity (intelligence)
Your voice. Your commands. AI ON!
With Generative AI voice assistance in Hyundai BAYON, interacting with technology takes on a whole new meaning. Simply say “Hey Hyundai” and explore a smarter way to connect with your drive.
https://x.com/HyundaiIndia/status/2109871755762516437
Hyundai Motor India has deployed SoundHound's real-time generative voice AI solution in the new BAYON SUV.
The system delivers continuous multi-turn conversations and 500+ embedded commands across 6 regional languages, operating both on-device and in the cloud.
https://x.com/SoundHound/status/2110724875736302016