TL;DR. SpaceXAI released Grok Voice 3.0, Fish Audio put out open weights for Drama 3 in 83 languages and OpenAI opened prompt-to-voice to all GPT-Live API customers. Speechmatics launched Oak for multilingual medical conversations. Fudan spinout Mosi Intelligence raised more than RMB 1 billion. Japan's actors union published its first rules for AI voices, a peer spoke in the House of Lords with an AI clone of her own voice, and Denver's council voted down a 311 voicebot.
New models and launches
SpaceXAI released Grok Voice 3.0 at $0.08 a minute (October 14). The speech-to-speech model is available in the Voice Agent API at the same price as Grok Voice 2.0 ($4.80 an hour). SpaceXAI reports a time to first audio of about 0.4 seconds and a score of 86.9 on Artificial Analysis' Speech-to-Speech Index. Starlink call centers use it to answer 20,000 customer calls a day in multiple languages, with roughly 70% resolved with no human involved, and it handles what the company calls millions of conversations a day in Tesla vehicles (vendor figures).
Source: SpaceXAI
Fish Audio released Drama 3, an open-weight TTS model in 83 languages (October 13). The backbone is a 35B-parameter mixture-of-experts with about 3B parameters active per token. Users direct delivery in plain language, can change emotion mid-sentence and edit single words without regenerating a whole line. Weights are on Hugging Face under the new Fish Audio Community License, free for commercial use below US$500,000 in annual revenue. The API lists at $50 per million characters, discounted to $10 until November 12. A technical report went up on arXiv the same day.
Sources: Fish Audio, arXiv
Speechmatics launched Oak, a medical speech recognition model (October 14). Oak transcribes clinical conversations that switch language mid-sentence, with no language set in advance, and supports 55 languages. In an internal Speechmatics study across English and 10 other languages, Oak-1 recalled 91.1% of medical terms, against an average of 74.3% for eight other providers including Google, Microsoft, AWS and Deepgram; in Spanish the figures were 96.3% and 86.8%. It runs in the cloud or self-hosted.
Source: GlobeNewswire
MBZUAI's Institute of Foundation Models and Inception42 launched Jais 3 with Jais Speech (October 15). The release pairs six open Arabic language models, from 900 million to 375 billion parameters, with Jais Speech, a recognition model for Modern Standard Arabic and dialects spoken across 19 Arab countries. Jais Speech handles switches into English or French, identifies the dialect and can add vowel marks. The developers report it had the lowest average word and character error rates of the models they compared across six public test sets.
Source: PR Newswire
Amazon's Bee launched two conversation-recording devices (October 15). Bee Frontier, a wearable for the wrist or clothing, costs $74.99 and is on sale now; Bee Eclipse, a magnetic phone attachment, will start at $99.99 when it ships. Both turn conversations into summaries and to-dos, connect to Alexa+ and Amazon.com, and process audio in real time before deleting it. Bee Pro costs $11.99 a month.
Sources: Amazon, TechCrunch
OpenAI opened prompt-to-voice to all GPT-Live API customers (October 16). Developers can create a custom voice from a text description in the API or the voice Playground and reuse it in GPT-Live sessions. Creating the voice is free; speech is billed at normal GPT-Live rates. OpenAI also added nine voices: three German, two French, one Hindi, two Japanese and one Korean. A day earlier it added native Decisions API integration to GPT-Live.
Source: OpenAI
Square launched a restaurant kiosk that takes voice orders (October 14). Square Kiosk has a voice layer that knows the seller's menu and handles questions, suggestions, modifications and substitutions, alongside a 21.5-inch touchscreen. It costs $1,999 with a counter stand plus a $99 monthly software fee per device, and requires a Square Plus or Premium plan.
Source: Square
Tether released QVAC TranscribePsy, a 254M-parameter on-device speech recognition model (October 14). It covers 25 European languages on laptops and 8 on phones. Tether reports 8.02% WER on the Open ASR Leaderboard and 6.58% on Multilingual LibriSpeech, and 10 minutes of audio transcribed in 1.65 seconds on an iPhone 17 Pro Max. Weights are open under Apache 2.0 for research and educational purposes, with GGUF and Core ML builds.
Source: Tether
Gnani AI launched Timbre v3.0, an expressive TTS model for Indian languages (October 14). It supports 10 languages, including Hindi, Bengali, Tamil, Telugu and Indian-accented English. Audio tags control whispers, laughter, shouting, pauses and pace, and the model can generate sounds such as coughs and sighs. Gnani says first audio arrives in under 200 milliseconds.
Source: Analytics India Magazine
SquadStack.ai released Arth V2, a streaming STT model for Indian phone sales calls (October 12). On SquadStack's benchmark of 863 real 8 kHz sales calls, Arth V2 scored an 8.85% semantic word error rate with 67 ms latency. SquadStack reports 18% fewer meaning-changing errors than Sarvam Saaras v4 and 28% fewer than OpenAI gpt-live-transcribe; Soniox v5 scored 8.77%. The model already handles 80% of the company's calls.
Source: SquadStack.ai
Also shipped this week:
Sarvam formally launched Bulbul V4 Flash: 11 Indian languages, under 250 ms to first audio, ₹20 per 10,000 characters (October 13). Source: Sarvam
Canonical released Myna, a local speech-to-text dictation app for Ubuntu Desktop, as a technology preview using Parakeet, Whisper or FunASR (October 12). Source: 9to5Linux
Google started rolling Gmail Live, voice search over the inbox, out to paid Workspace business editions on mobile, US English only (October 15). Source: Google Workspace Updates
Xiaomi open-sourced MiDashengLM-Spatial, an audio-language model that locates where sounds come from, under Apache 2.0 (reported October 16; paper October 8). Sources: TechNode, arXiv
Druid AI added purpose-tuned voice models for regulated industries to its agent platform (October 14). Source: Business Wire
Hedy AI launched Hedy for Teams with on-device transcription and an EU data region, at $99.99 per seat per year (October 13). Source: GlobeNewswire via Yahoo Finance
Mistral published Voxtral Mini 4B Realtime Arabic on Hugging Face, an open-weight (Apache 2.0) streaming speech-to-text model for Modern Standard Arabic and 15 dialects, at 8.82% average CER with a 480 ms delay; there was no announcement (October 8, missed last week). Source: Hugging Face
Embedding Gemma Transcribe, an independent open-source project, turned Google's EmbeddingGemma 2 audio encoder into an offline Mac dictation app; its developer reports about 5% WER and an hour of audio in about 13 seconds on a base M3 MacBook Air (October 14). Source: Embedding Gemma Transcribe
Infrastructure and tooling
Speechmatics brought Melia 1 to real-time transcription (October 13). The multilingual model detects the spoken language across 56 languages without a preset, handles word-level code-switching mid-stream and labels every word with its language. It runs in the US, EU and Australia regions. On October 16 it also updated the Melia 1 batch model, with a 1.6% word error rate improvement in English.
Source: Speechmatics changelog
ElevenLabs added audio effects to its TTS API and turn prediction to real-time STT (October 12). Text to Speech and Text to Dialogue requests accept an optional audio_effects object for a filter preset, perceived distance, reverb, a background noise bed and stereo panning, at no extra cost. The real-time Speech to Text WebSocket adds a turn_prediction commit strategy that commits a segment as soon as the model predicts the speaker has finished, and partial transcripts carry a turn_probability score.
Source: ElevenLabs changelog
LiveKit Agents 1.8.8 runs OpenAI GPT-Live through LiveKit Inference (October 15). Developers can call openai/gpt-live-1 with LiveKit credentials instead of a separate OpenAI key. The release adds model strings for Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, swaps SpaceXAI's Grok Voice 3.0 in for Think Fast 2.0, and adds a background decision-model plugin. Version 1.8.7 (October 14) added an end-to-end latency metric for realtime replies, measured from the last detected user speech to first assistant playback.
Sources: GitHub, 1.8.8, GitHub, 1.8.7
Telnyx put video avatars on its voice AI assistants (October 13). Six preset avatars, rendered on Telnyx infrastructure, are in beta and switched on per conversation. They work only with Telnyx voices, not third-party TTS such as ElevenLabs, and fall back to voice-only on errors. Avatars are included in call pricing through October and cost $0.08 a minute from November 1.
Source: Telnyx
Also in the changelogs:
Vapi added percentage traffic splitting across assistant versions, and GPT-Live calls now accept keypad (DTMF) input (week of October 12). Source: Vapi
Sarvam put Instant Voice Cloning (beta) in its Python and TypeScript SDKs (October 12), set Bulbul V3's end of life in Voice Agents for November 10 (October 14) and moved call recordings to stereo, one speaker per channel (October 16). Source: Sarvam changelog
Deepgram SDKs can now reconfigure an active Nova-3 stream, including keyterms, without reconnecting (October 12). Source: Deepgram changelog
NineNineSix released Tyger, an open-weight turn-taking model for full-duplex voice agents built on NVIDIA's Nemotron 3.5 ASR Streaming 0.6B (October 12). Source: Hugging Face
Twilio Voice Integrity, which registers business numbers with US analytics vendors to cut spam labelling, now covers all three major US networks (October 16). Source: Twilio
Telnyx launched Programmable Voice in Mumbai with local SIP connectivity (October 16). Source: Telnyx
Microsoft Azure Speech released a multilingual speech-to-text model for 13 Indic locales in Fast Transcription and Voice Live, about 10% better than the prior generation (October 5, missed last week). Source: Microsoft Learn
The money
A Fudan University spinout raised over RMB 1 billion for real-time voice and video models (October 13). Mosi Intelligence, founded by Fudan professor Qi Xipeng and led by CEO Li Shimin, a core author of SpeechGPT, closed a Pre-A round of more than RMB 1 billion led by Huakong Fund. Co-investors include China Development Bank Science and Technology Innovation and the Shanghai Science and Technology Innovation Fund. The company builds MOSS Live, a real-time interactive system that understands and generates voice and video, and says it released more than 20 model versions in six months.
Source: 36Kr
A former Plaud executive raised $26 million for an always-listening AI app (October 12). Peter Mo's Ambi has raised $26 million across multiple rounds, led by Luminous Ventures with participation from device maker Huaqin. The app runs on iPhone, Mac and Apple Watch, listens in the background and produces meeting notes and suggestions. Ambi plans its own device with a screen, speaker, camera and microphones, and has more than 50 staff in San Francisco, Beijing, Shanghai and Shenzhen.
Source: TechCrunch
Othello raised an $8 million seed round for live sales call coaching (October 14). Investors are a16z speedrun, Owl Capital and Long Journey Ventures, plus angels including Hugging Face co-founder Thomas Wolf. Its in-house model, Hamlet, gives reps suggestions in under 500 milliseconds while the buyer is still on the call. The company says it is deployed at more than 50 large enterprises, including Microsoft, Capital One and Lyft.
Source: PR Newswire
MEDFAR bought AlloMia to answer clinic phones with voice agents (October 14). The Montreal company behind the MYLE medical records platform did not disclose terms. AlloMia's multilingual agents book appointments and route patient calls; in the most recently onboarded pilot clinics, about 50% of calls are handled end to end without a transfer to reception. It is MEDFAR's second AI acquisition in under two years.
Source: MEDFAR via Newswire.ca
More deals:
Leadrat acquired AI calling startup Calllive.ai for an undisclosed sum, to add voice agents to its real estate CRM in Hinglish and other Indian languages (October 12). Calllive.ai reports more than 5 lakh calls; the product will be rebranded CallingCat. Source: CXOToday
Dapta raised a $6 million seed round led by Chingona Ventures for Dapti, a sales agent that answers leads by phone, text and WhatsApp in English or Spanish (October 13). Source: Dapta
VoicePatrol, a Romanian-founded startup that moderates toxicity in multiplayer game voice chat, raised a EUR 3 million seed round led by GapMinder VC (October 12). Source: The Recursive
Vocca raised a $20 million Series A led by Norrsken VC for AI phone agents for medical practices (October 5, missed last week). It reports 15,000 practitioners, more than 1 million calls a month and 70% of conversations resolved without a human. Source: HIT Consultant
Law, safety, and consent
Japan's actors union published its first rules for AI voice use (October 15). The Japan Actors Union, which represents about 2,600 actors and voice actors, requires the performer's prior consent at each stage: training, generation and public use. The voluntary guidelines also call for payment for AI-generated output, separate consent and pay for AI dubbing into other languages, and the permission of heirs for the voices of deceased performers. The union says Japan has no law that protects the voice itself and will lobby for one. The same day, its vice-chair Masako Nozawa, the voice of Goku, told NHK that AI videos using her voice without permission are "like theft".
Sources: Japan Actors Union, Tbreak
A peer spoke in the House of Lords with an AI clone of her own voice (October 13). Baroness Campbell of Surbiton, whose disability increasingly affects her speech, used an ElevenLabs voice built from recordings made before 2012 to speak in the second reading of the Health Bill. Disability News Service reports it was the first time an MP or peer has used an AI voice generator to deliver a speech in Parliament.
Source: Disability News Service
Denver's City Council voted down a 311 voicebot (October 13). The council rejected 9-3 a five-year, $3.9 million extension of the city's contract with Citibot, which runs the "Sunny" chatbot and would have added a voicebot for "low complexity" 311 calls. Councilman Paul Kashmann said: "I think artificial intelligence is a nightmare unfolding before our eyes." The mayor's office says it will look for a way to keep the service running.
Source: The Denver Post via GovTech
Mexico's lower house passed a takedown duty for AI-altered voices (October 7, missed last week). The Chamber of Deputies approved changes to the Federal Copyright Law 325-114, with one abstention. Platforms must remove works, images, voices and recordings "modified or altered without authorisation through artificial intelligence or other technological tools" when notified, and cut off repeat infringers. Fines go up to about 4.69 million pesos. Critics call it the "anti-meme law"; it goes to the Senate.
Sources: Voice Over Herald, La Razón
Lithuanian lawmakers proposed a legal right to one's voice (October 9, missed last week). A Civil Code amendment led by MP Vytautas Kernagis would require consent to reproduce, publish or use a person's voice, including AI-generated imitations that identify them. Consent to record a voice would not cover building a synthetic one. Parody and satire would be allowed if they do not mislead people about authenticity.
Senior managers in Singapore were the most likely to obey a cloned CEO voice (October 13). In a KnowBe4 survey of 500 office workers by Censuswide, 36% said they cannot confidently tell a genuine CEO message from an AI-generated one. Senior managers were the most confident at spotting fakes (78%) but 35% would comply with a cloned CEO voice before verifying, against 16% of clerical workers. KnowBe4 sells security training.
Source: KnowBe4
Also this week:
Claudio Serrano, the Spanish voice of Batman, said an AI platform hosted an unauthorised clone of his name, image and voice, and asked him to complete a voice captcha to file his complaint (reported October 12). The platform is not named and has not responded. Source: Voice Over Herald
Inka Magnaye, a Filipino voice talent, said her mother, veteran voice artist Lindy Pellicer, lost a long-held voice-over job at a major brand to AI (October 8, missed last week). The brand is not named. Source: Interaksyon
In the wild: products, enterprise, and culture
CMA CGM will take its HappyRobot sales voice agent to 20 countries (October 14). In first campaigns in the UK and Senegal, the agent, called Eva, contacted more than 4,000 prospective customers who had no dedicated sales rep, and about a third went on to active discussions about shipping routes. CMA CGM plans to deploy Eva in 20 countries by the end of 2026. Separate HappyRobot agents for cargo release enquiries resolve more than 46% of those requests without staff, and finance agents chase unpaid invoices. CMA CGM handles more than 54 million customer requests a year.
Sources: Hellenic Shipping News (release text), The Loadstar
NatWest cut misheard words by 20% after tuning Nvidia Nemotron speech recognition to UK accents (October 14). 90 NatWest staff recorded about 3,000 banking questions in accents including Glaswegian, Northern Irish, Welsh, Yorkshire and Mancunian. Nemotron ASR turns speech into text, and NatWest's own small language models interpret it. The speech recognition runs in Spending Insights, a voice and text tool now available to 75% of Royal Bank of Scotland customers, with wider NatWest Group rollout planned. One example from testing: a Scottish speaker's "last quarter" was transcribed as "last water".
Source: NatWest Group
Samsara made Agent Studio generally available, and its voice agents now outrun chat (October 13). Agent Studio is a no-code builder for AI agents in fleet and field operations, now open to all customers. Samsara says agent adoption has grown 7x since June, voice agents passed chat in interaction volume within four weeks of launch and have led every week since, and about 75% of organizations using voice agents have increased weekly usage. These are vendor figures.
Source: Samsara
Yum Brands says Taco Bell's drive-thru voice AI is improving staff retention (October 14). Chief legal officer Erika Burkhardt told the Fortune Most Powerful Women Summit: "AI is actually increasing retention among our team members in the restaurant because it's making their jobs easier." Omilia voice AI takes orders at nearly 900 US Taco Bell drive-thrus. Yum gave no retention figures.
Source: Fortune
Cerence and Geely took their voice partnership from cars to a humanoid robot (October 15). At the Paris Motor Show, the two companies showed a Geely humanoid robot running Cerence's speech recognition and conversational AI, which handles French, English and Mandarin without the user switching modes. It is the first international showing of the robotics collaboration.
Source: Cerence
Rivian plans to let third-party AI agents into its cars (October 15). At TechCrunch Disrupt, CEO RJ Scaringe said Rivian's software will work with agents from companies such as Anthropic, OpenAI and Meta alongside its in-house assistant, with simple requests handled on board and complex ones sent to the cloud. Agents will not control driving functions such as braking or steering. Rivian gave no date and pointed to its AI & Autonomy Day in December.
Source: TechCrunch
A Jabra survey links audio hardware gaps to voice AI use at work (October 13 to 15). In Jabra's survey of 1,000 UK knowledge workers, 31% named at least one hardware barrier to using AI effectively, and 13% said a lack of reliable headsets and microphones stops them using voice AI tools such as meeting transcription. Regular voice AI users reported less stress (20% against 26% of non-users). Jabra sells headsets.
Source: AV Magazine
More voice agents went live: (October 8 to 16)
Flair Labs and real estate platform CINC launched CINC Call AI, which calls new leads, qualifies them and hands them to agents. CINC is used by more than 50,000 US real estate agents (October 15). Source: EIN Presswire
Aiello certified its hotel call agent, voice assistant and guest web app with Shiji Infrasys POS, so orders go straight to the restaurant system. Aiello says Holiday Inn & Suites Saigon Airport handles more than 450 food and beverage orders a month through it (October 13). Source: PR Newswire
BYD and HONOR linked HONOR's YOYO phone assistant to BYD's DiLink in-car agent over the A2A protocol, starting with voice commands to find and prepare the car (October 12). Source: Sina Tech
NTT Docomo Business published results from CONN, an assistant built on ElevenLabs agents that gives older residents of Aizuwakamatsu, Fukushima, local information and conversation, in trials since autumn 2025 (October 16). Source: ElevenLabs
ElevenLabs opened a Singapore office as its Southeast Asia hub, naming customers including Funding Societies, Atome and Boost Bank (October 8, missed last week). Source: ElevenLabs
Papers this week
General speech models misidentify Indian languages on short clips (arXiv, October 14). Whisper-style models have tokens for only 15 of the 23 languages tested, and on one-second inputs Gemini 3.5 Flash reaches 0.52 macro-F1 and GPT-audio 0.14; the dedicated 115M-parameter Anuvani-LID, trained on 54,643 hours, reaches 0.914 mean macro-F1 across 37 conditions. Source: arXiv
Zero-shot diarization across all 22 scheduled Indian languages (arXiv, October 14). Anuvani-Diarization reaches a 16.78% duration-weighted DER on 1,164 IndicDiarBench recordings, and the benchmark's annotation convention alone shifts DER by 1.9 to 4.1 points depending on the system. Source: arXiv
SITAR scores ASR errors by how much they distort meaning, without a runtime LLM judge (arXiv, October 15). A lightweight BERT-family regressor scores each reference-hypothesis pair in one forward pass, trained on targets built with limited offline LLM support; the authors say code, weights and data will be released after internal review. Source: arXiv
Robots that act before the sentence ends (arXiv, October 14). EarlyCommit scores scene objects after every recognized word and commits correctly 1.6 to 2 times as often as keyword matching on unseen words, is right at least 92% of the time when it commits at a 0.9 threshold, and needs 31 to 35 ms per word on consumer GPUs. Source: arXiv
Longer voice prompts carry more prosody, up to about 10 minutes (arXiv, October 14). LP-TTS-Bench covers 37 speakers, 63.95 prompt hours and 2,728 held-out test utterances; prosody gains over a short-prompt baseline grow with prompt length up to roughly 10 minutes, then saturate. Source: arXiv
Why streaming TTS degrades past its training length (arXiv, October 14). Analysing VoXtream2, the authors find pronunciation fails near the end of the trained position range and speaker similarity declines throughout; two training-free attention changes keep both stable in long-form generation. Source: arXiv
Whisper is not a reliable intelligibility judge for Kazakh TTS (Algorithms, October 15). On 138 test sentences VITS reached 1.94% and 2.08% CER against 8.79% for Tacotron 2 when scored by a Kazakh-specific ASR model, while multilingual Whisper reported about ten times higher error rates on the same audio. Source: MDPI
Evaluation protocol moves emotion recognition scores more than model choice (DÜMF Mühendislik Dergisi, October 17). Across 12,162 recordings from 121 speakers, random splits inflated accuracy by 14.61 points on average, more than the 11.24-point gap between the best and worst of seven models; cross-dataset, the best model reached 29.3% balanced accuracy on six classes. Source: DÜMF
Adapting Cohere Transcribe for Hassaniya Arabic (preprint, October 15). Fine-tuning on 37.42 clip-hours cut CER from 40.5% to 8.1% on a media test set and from 42.0% to 18.5% on the Casablanca split, with 6.1 times real-time throughput on CPU with INT8 weights. Self-published. Source: Zenodo
"Blind" audio benchmark scores do not show whether models listen (preprint, October 15). Re-analysing a 2026 audit in which models kept 60 to 72% of accuracy without audio, correcting for guessing lowers the retained share to 29 to 45%. Self-published. Source: Zenodo
Customers judge voice bots as one overall impression (JTAER, October 13). A survey of 295 voice bot users in Israel across eight service categories found one general factor accounts for most shared variance across quality facets, and interaction complexity had no significant effect. Source: MDPI
Previous edition: Voice AI News, Week 41, October 5 to 11, 2026