TL;DR. ElevenLabs released Scribe v3, Salesforce open-sourced SOV-Transcribe-1 and Reson8 shipped Rumble-2, while Fish Audio, Bland and Retell launched new models for voice agents. OpenAI will retire ChatGPT's older voice modes on December 11, Gemini Live reached Apple CarPlay and Meta's Muse started calling businesses. Twilio made Smart Turn Detection generally available. Retell says it passed a $100 million run rate, and a Warsaw court ruled that an AI clone of a voice actor's voice was a deepfake.
Community
Voice AI Day is back on December 10, with free meetups in 11 cities on the same evening (announced November 16). Voice AI Space is running the second edition of its global community meetup, after the first on August 20. In-person events take place in Bengaluru, Dubai, Gdańsk, Istanbul, London, Melbourne, Mumbai, New York, Paris, San Francisco and Toronto, each starting in the evening, local time. Attendance is free and registration for each city is on Luma.
Source: Voice AI Day
New models and launches
ElevenLabs released Scribe v3 and Scribe v3 Live, new speech-to-text models for recorded and real-time audio (November 10). Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages on an internal version of FLEURS. Scribe v3 Live has a 31% lower word error rate than Scribe v2 Realtime, according to the launch post, and adds a mode for speakers who switch languages mid-sentence. Both are in the ElevenLabs API; the post gives no pricing.
Sources: ElevenLabs, ElevenLabs changelog
Salesforce released SOV-Transcribe-1, an open-weight speech-to-text model under Apache 2.0 (November 12). The model pairs a Whisper large-v3 audio encoder with a Gemma 4 language model from Google and transcribes recorded audio in seven languages. Salesforce reports a 2.28% word error rate on Artificial Analysis's benchmark, first among open-weight models as of November 12 (vendor figure).
Sources: Hugging Face blog (Salesforce), Model card
Reson8 released Rumble-2, its most accurate speech-to-text model, a week after Rumble-2-Flash (November 10). Rumble-2 runs in batch and streaming, covers 50+ languages and takes custom vocabulary without fine-tuning. On Cekura's independent benchmark it reports a 2.05% word error rate, first among the models listed, ahead of Fish Audio (2.21%) and AssemblyAI Universal 3.6 Pro (2.71%). The post gives no pricing.
Sources: Reson8, Reson8 blog index
Zyphra open-sourced ZONOS3-Mini, a text-to-speech model with about 100 million active parameters (November 9). It is a sparse mixture-of-experts model with 800 million total parameters, released under Apache 2.0 with voice cloning from a reference clip. Zyphra reports time to first audio under 150 ms in a browser on an M3 MacBook Air (vendor figure). Weights are on Hugging Face.
Sources: Zyphra, Hugging Face
Fish Audio launched Live 3, a full-duplex voice model that keeps listening while it speaks (November 10). Fish Audio says the model was trained on real human conversations to decide when to acknowledge, wait, respond or stop, and can decide to respond in as little as 186 milliseconds (vendor figure, no test conditions given). It supports English only at launch. No pricing or benchmarks were published.
Source: Fish Audio
DeepL introduced Aure, a voice model for real-time spoken translation that keeps the speaker's own voice (November 11). Aure adjusts the translated speech to the speaker's tone, emotion, emphasis and pace, and preserves their voice identity. It applies automatically across DeepL Voice for meetings and in-person conversations and the DeepL API for Voice, in eight languages at launch.
Source: DeepL
Retell launched Retell LLM (Tell-1), its own language model for voice agents, which starts speaking before it finishes reasoning (November 13). The model writes the first sentence of a reply, reasons while that sentence plays, then continues. Retell reports a 72 ms median time to the first meaningful chunk (vendor figure, model time only). It is priced at $0.035 per minute and becomes selectable on November 19.
Source: Retell
Google added offline speech recognition and text-to-speech to its AI Edge Gallery app and developer tools (November 10). Speech recognition covers 100+ languages and text-to-speech offers 8 voices across 21 locales. On Android it uses ML Kit's new GenAI speech APIs, built into the OS; on other platforms it runs open-weight models such as Parakeet-TDT 0.6B, Kokoro-82M and Gradium's Phonon 31M. The Gallery code is open source.
Source: Google Developers Blog
Bland opened a research preview of Tofu-1, a full-duplex speech-to-speech model for phone calls (November 13). Tofu-1 processes audio directly, can be interrupted or corrected mid-sentence, and keeps the conversation going while it completes tasks. Access is limited to a small number of enterprises through a request form; Bland has not published model size, latency, pricing or benchmarks.
Sources: Bland, Bland on X
Also shipped this week:
ElevenLabs opened free early access to ElevenAI, an app that takes meeting notes, summaries and action items and includes dictation, announced at its New York summit (November 11). Sources: ElevenLabs, TechCrunch
Otter added 13 transcription languages, with 14 more and a language-switching mode marked "coming soon", for 31 languages in total once the rollout is done (November 10). Source: Otter
Altius Link, a Japanese contact center operator, began offering AI Voice Service, an AI phone operator for product inquiries, reservations and fault reports; it names no customers yet (November 11). Source: Altius Link
Chariot, an Indian AI lab, launched Chariot-1-TTS-Multilingual, a text-to-speech model for 11 Indian languages and English with more than 50 voices and voice cloning from a few seconds of audio (November 12). Source: CIOL
Infrastructure and tooling
Twilio made Smart Turn Detection generally available in ConversationRelay (November 13). An in-house model decides in real time whether a caller's pause ends the turn, instead of relying on silence timers. It works with all supported speech-to-text providers except Deepgram Flux. Twilio published no latency, accuracy or pricing figures.
Source: Twilio changelog
ElevenLabs opened a marketplace for partner-built voice apps and launched a partner program at its New York summit (November 11). The ElevenLabs Marketplace lists industry-specific agents and voice apps built on ElevenLabs, with Salesforce, IBM, NiCE, Twilio and Zendesk as launch partners. Enterprises can browse and contact partners now. The new Partner Program adds certifications, deal registration and co-selling.
Sources: ElevenLabs, Marketplace, ElevenLabs, Partner Program, ElevenLabs, NYC summit
Pipecat 1.16.0 added a Zoom AI Voice Agent service and audio classifiers that tell the LLM how the user sounded (November 12). The release adds a service for speech-to-speech conversations through the Zoom AI Voice Agent, which handles turn detection and barge-in itself. An alpha feature queries classifiers on each user turn, including the new open Pipecat Voice Reader model (tone of voice, laughter, sarcasm, speech rate) and Hume's Audio Decisions API.
Source: GitHub, Pipecat v1.16.0
LiveKit Agents 1.8.12 added recovery for answers GPT-Live hears but does not act on, plus new Google, Perplexity and AssemblyAI models (November 13). With a new setting, a detector checks for caller answers that GPT-Live transcribed but never delegated to the backend and starts the delegation after a quiet grace period. New models include Gemini 4 Carbon and Gemini 4 Hydrogen, Perplexity Decider, AssemblyAI streaming STT, and Lokutor STT and TTS in nine languages.
Source: GitHub, LiveKit Agents 1.8.12
Microsoft made real-time multichannel speech-to-text generally available in Azure Speech (November 9). A single real-time session transcribes up to two channels independently, such as agent and customer on a contact center call, and each channel can be recognized in its own language. It is priced the same as single-channel real-time transcription.
Source: Microsoft Tech Community
Also in the changelogs:
Vapi added Reson8 and Soniox stt-rt-v6 as transcribers (week of November 9). Source: Vapi changelog
Speechmatics added speaker diarization to real-time transcription with its Melia 1 and Oak 1 models (November 12). Source: Speechmatics changelog
Telnyx can now silence card, personal and health data in call recordings before they are stored, for English speech (November 13). Source: Telnyx release notes
The money
Primordial Labs raised an $18 million Series A for Anura, voice command software for military systems (November 10). Ballistic Ventures led the round, with existing investors Squadra Ventures and Lockheed Martin Ventures. Anura lets crews direct onboard sensors, communications and connected uncrewed systems by voice instead of through manual controls. The New Haven company says licenses are being delivered to the Army and SOCOM.
Sources: Primordial Labs (PR Newswire), Axios
Retell AI says it passed $100 million in annualized run rate, with $4.6 million raised in total (November 12). The voice agent platform says it reached the figure in under three years, with more than 4,000 customers running over 100 million calls a month. These are company figures.
Source: Retell AI
Dialpad and Ada agreed to combine into one company for AI customer service (November 4, missed last week). Dialpad brings its voice and communications platform and Ada its AI customer service agents; together they serve more than 50,000 companies. Craig Walker will be CEO of the combined company. Financial terms were not disclosed.
Source: Dialpad
NiCE is cutting about 5% of its global workforce as third-quarter revenue rose 8% (November 4, missed last week). The contact center software company reported revenue of $790.7 million and said its AI annual recurring revenue reached $392 million, up 46%. NiCE expects $15 million to $20 million in pre-tax restructuring charges.
Source: NiCE press release (via Seeking Alpha)
More deals:
Helix, which uses a four-second voice challenge on the user's device to check that they are a unique human and to estimate their age, raised $6.2 million led by Chicago Ventures. Siri co-founder Dag Kittlaus is backing the company and joins as an adviser (November 11). Source: TechCrunch
PromptPath, an Austin conversation assist platform that gives car dealership staff real-time prompts during customer conversations, raised a $12 million Series A led by FM Capital and Armory Square Ventures, with Automotive Ventures, Holman Ventures and dealer groups (November 10). Source: PR Newswire
Modulate added Twilio Ventures and Kayak co-founder Paul English as investors, amount not disclosed, following its $25 million round in September. It also said its Velma platform now flags a synthetic voice in a median 1.6 seconds, down from 2.25 seconds (November 10). Source: Modulate
Worker.ai, a Raleigh small business software company backed by Veeam founder Ratmir Timashev, announced it has acquired DeltaGen, a Techstars-backed startup whose real-time voice technology captures decisions and next steps from live conversations. No terms were given (November 12). Source: PR Newswire
PartnerOne, an enterprise software investment group, acquired 3CLogic, a voice AI and contact center platform with its own communications infrastructure. No terms were given (November 3, missed last week). Source: PR Newswire
Law, safety, and consent
Poland: a Warsaw court ruled that an ad using an AI clone of a voice-over artist's voice was a deepfake (November 13). The Warsaw Regional Court found that JFC Polska, a maker of home sewage treatment tanks, used an AI-generated copy of lektor Jarosław Łukomski's voice in adverts he never recorded or approved. The court ordered the company to remove the material, publish apologies and pay 30,000 zlotys in compensation and more than 17,000 zlotys for losses. The judge described it as the first case of its kind in Poland. The ruling is not final.
Sources: RMF24, Voiceover Herald
UK: Alexa, Google Assistant and Siri were designated as regulated radio services, including their AI versions (November 12). The government confirmed that Amazon's Alexa (including Alexa+), Google Assistant (including Gemini) and Apple's Siri are designated as Radio Selection Services under the Media Act 2024. Once the regime is fully in force, the assistants must play a UK station's online stream when a listener asks for it by voice, without inserting their own ads before or during it or charging stations for access.
Sources: UK Parliament written statement, GOV.UK consultation
Also this week:
US: a Washington consumer filed a proposed class action under the TCPA and Washington state law against Allstate, the Hardin Agency and lead platform InboundCD, alleging unsolicited "AI voice outbound calls" pitching insurance. These are allegations (reported November 13). Source: TCPAWorld
In the wild: products, enterprise, and culture
Google brought the Gemini app and Gemini Live to Apple CarPlay (November 11). Drivers can hold hands-free conversations with Gemini Live from the car display, ask for places in the area they are heading to, catch up on unread Gmail and dictate replies. Google says the app is rolling out globally, starting November 11, to all Gemini users on iOS.
Source: Google
OpenAI will retire Advanced and Standard voice in ChatGPT on December 11 (November 10). After December 11, the two older modes disappear from the voice picker and users are moved to GPT-Live, the full-duplex model introduced in July. OpenAI says "virtually all" voice users now use GPT-Live. Live video and screen sharing will keep running on Advanced voice until GPT-Live supports them.
Source: OpenAI Help Center
Meta's Muse can now place phone calls to US businesses for paid subscribers (November 12). Meta chief AI officer Alexandr Wang listed insurance, cable, DMV and dinner reservations as examples and said the feature is "out now to paid subs in the US." TechCrunch reports paid plans start at $20 a month. On November 9, Muse also arrived on Apple Watch and Wear OS.
Sources: Alexandr Wang on X, TechCrunch, TechCrunch, Muse on watches
Apple HomeView, a $299 home hub built around Siri AI, went on sale (November 13). HomeView answers questions with Siri AI, can take action across apps, and recognizes up to six people by face to show personal content. Siri AI on HomeView is in beta, in English only, in eight countries including the US and the UK. HomeView is not sold in the EU because Siri AI is unavailable in iOS there.
Source: Apple Newsroom
Spotify gave its AI DJ nine new voices and an option to turn off commentary (November 10). Premium users can now pick from seven new English voices and two new Spanish voices, alongside the original DJ X. A new setting turns off talk between songs. Spotify says 80% of DJ listeners come back week after week.
Source: Spotify Newsroom
Customers Bank, Westgate Resorts and Yahoo described voice agents built on ElevenLabs (November 11). At ElevenLabs' New York summit, ElevenLabs said Customers Bank uses several of its agents, including a 24/7 voice and chat support agent. Westgate Resorts launched "Sarah", a front-desk voice agent, at its 2,100-room Orlando resort. Yahoo added real-time voice conversations to its Scout answer tool. Launch dates were not given; all details come from ElevenLabs' post.
Source: ElevenLabs
More voice agents went live:
Poland's national e-Health platform uses a voice agent built on ElevenLabs models and NVIDIA infrastructure that calls landline-only patients to remind them of appointments and help them reschedule; the agent runs on-premises (November 11). Source: ElevenLabs
Avis Budget Group now runs all its contact centers on Amazon Connect Customer, with an AI-powered IVR that understands natural speech; AI voice assistants for sales and service requests are still in testing (November 9). Source: AWS
Papers this week
Commercial voice agents lose track of tasks when users change requests or interrupt mid-answer (arXiv, November 9). Joi-Bench-Task, from Tencent's Platform and Content Group, ran GPT Live 1.0, GPT Realtime 1.5, GPT Realtime 2, Grok Voice 3.0 and Gemini 3.8 Live through 140 simulated test cases each (700 conversations) in which a scripted user changes the request or interrupts while tools run or results are being read out. When requests changed, the best result completion rate was 55% (GPT Realtime 1.5) and the lowest 15% (Gemini 3.8 Live); after an interruption, GPT Live 1.0 omitted 3.5% of the information it still owed the user, while the other four omitted 23% to 79%. Source: arXiv
Dictation ASR systems with the same word error rate fail very differently on whispered and mixed-language speech (arXiv, November 12). A Lemon team streamed live dictations to eleven ASR systems and sorted every error by type: Zoom Scribe and OpenAI's GPT-Live-Transcribe both score about 7.3% to 7.4% WER on quiet English, but 34.7% and 8.8% on whispered English, and Zoom and both AssemblyAI models dropped 26% to 28% of whispered words. With only English selected, Gemini, Deepgram and a self-hosted Whisper lost the non-English part of two-language recordings in 86% to 99% of cases, and putting Whisper behind a Silero voice activity detector that drops audio it judges silent raised whispered WER from 10.1% to 51.4%. Most English results come from one read speaker; recordings and transcripts are released. Source: arXiv
Observe.AI finds another AI answered 6.9% of calls in a 1,000-call sample of its outbound voice AI calls (arXiv, November 13). In 1,000 sampled outbound calls, a live person answered 48.1%, voicemail 44.6%, an AI 6.9% and an IVR 0.4%, so the team built FEND, which labels the far end as human, voicemail, IVR or AI callee within 20 seconds instead of the single human-or-machine bit of standard answering machine detection. Speech timing separated AI callees from humans at 0.952 ROC-AUC because synthesized speech leaves shorter gaps between words; on production traffic FEND reached F1 0.979 for live person versus automated, labeling 2.9% of live people as automated (company figures). Source: arXiv
Listeners rank TTS systems differently for "sounds human" than for "would rather talk to" (arXiv, November 11). Fluid Concepts Research tested six commercial systems (Cartesia Sonic 3.6, Deepgram Aura-2, ElevenLabs v4 Turbo, Gemini 2.5 Flash TTS, OpenAI gpt-4o-mini-tts and Rime Coda) on customer-service turns with 654 listeners. Against the original actor recordings, systems were chosen 19% of the time for human-likeness but 36% for conversational preference, and Cartesia reached 49% on preference; variation between voices of one system was about three-quarters of the variation between systems. The UTMOS and UTMOSv2 naturalness scores predicted lower human-likeness, and none of the three automatic scores tested predicted preference. Source: arXiv
NVIDIA's Nemotron 3 Diarization doubles streaming speaker capacity to eight while running faster (arXiv, November 9). The 99M-parameter successor to Streaming Sortformer replaces the FastConformer stack with a pure Transformer encoder, cuts training memory by a factor of 2.8 and gives up to 6.4 times higher inference throughput. At 1.04-second latency its diarization error rate is 13.18% on DIHARD III and 10.29% on CALLHOME, against 19.60% and 11.31% for Sortformer-v2.1. Weights are on Hugging Face under the OpenMDW-1.1 license, with training code open source. Source: arXiv
Previous edition: Voice AI News, Week 45, November 2 to 8, 2026