Gemini 3.8 Live, Grok Transcribe 2.0 and much more this week

Voice AI weekly digest

Events 🎤

• Vapi hosts “Enterprise voice AI after hours” at HumanX Amsterdam (Sep 23, Amsterdam, LUMA)

• AssemblyAI runs “Build Night: Create Your Own Dictation App” (Sep 24, San Francisco, LUMA)

• Google DeepMind hosts “Gemini Audio | At Night” (Sep 24, San Francisco, DeepMind)

Top Updates 💪

• Google launches Gemini 3.8 Live, voice model that reason while speaking and tops the Artificial Analysis Speech-to-Speech index. (Google)

• xAI ships Grok Voice Transcribe 2.0, calling it twice as accurate as v1 and ranking first among streaming models on the Artificial Analysis leaderboard. (xAI)

• ServiceNow integrates Krisp’s Voice Isolation to power its Voice Agents (LinkedIn)

• Speechmatics launches Agent STT, a model built to catch the high-consequence errors, like a wrong digit or missed negation, that derail voice agents. (Business Insider)

• Apple launches Siri AI, a rebuilt assistant with personal context, onscreen awareness, and systemwide app actions, rolling out in beta in English. (Apple)

• StepFun ships StepAudio 3 Realtime, a think-while-speaking duplex model that posts 98.9 on the Full-Duplex voice benchmark. (AI Weekly)

• Wispr Flow debuts Canto, its first speech model, cutting noisy real-world error rates from over 30% to single digits (Wispr Flow)

• DeepL Voice now preserves your voice, carrying tone, rhythm, and intonation across real-time translation in 30+ languages on Zoom, Teams, and Meet. (PR Newswire)

• Microsoft adds real-time voice agents to Copilot Studio, letting organizations build speech-to-speech agents first via Dynamics 365 Contact Center. (Microsoft)

• Superhuman acquires AI notetaker Fathom, folding bot-free meeting capture into its email, calendar, and agent suite. (citybiz)

• Treble raises $18M to expand its voice simulation platform, generating synthetic acoustic data for model training, robotics, and consumer hardware. (TechCrunch)

• Google pushes AI for every language, extending its speech stack toward the 1,000 most-spoken languages via cross-lingual transfer. (Google)

• Radisys launches the V.AI ecosystem, bundling ElevenLabs, Hiya, Speechmatics, and others so telcos can deploy voice AI inside their own networks. (Korea Newswire)

• Zoom pivots from video calls to agentic work, positioning ZoomMate to search across systems and finish follow-up tasks, not just summarize. (WebProNews)

• WhatsApp tests speech-to-text messaging, letting users dictate a message and send it as text, processed on-device and offline. (The News)

• Goodlord warns AI voice spoofing is gaming tenant referencing, with fraud rings using real-time voice tools to bypass reference callbacks. (Goodlord)

• Research shows scammers clone local accents, exploiting familiarity to lower listeners’ guard as AI voice scams surged over 1,200%. (Android Headlines)

• Google DeepMind argues AGI will be spoken, not typed, making the case for one natively multimodal speech-to-speech model over cascaded pipelines. (BigGo)

• FineVoice launches Aunio, an AI audio production agent that coordinates voices, music, and sound effects into publish-ready audio end-to-end. (PR Newswire)

• RingCentral rolls out AIR Pro for Healthcare, an agentic voice AI with prebuilt patient workflows and native integration to 100+ EHR systems. (RingCentral)

• ImpactFactory.ai ships version 1.5, adding a full-duplex conversational agent that guides users through marketing tasks. (EIN Presswire)

• Fierce argues telcos must embrace voice AI or get left behind, as Alianza’s Crux pitches operators a path beyond being a dumb pipe. (Fierce Network)

---

Engineering Corner 😎

• Google details building real-time voice apps with Gemini Audio, its developer guide to the new Live and Extended Thinking models. (Google)

• A benchmark comparison of speech-to-text APIs finds Meta Muse leading streaming WER at 3.1% while undercutting Google’s pricing. (Tech Insider)

• Nari Labs tops Coval’s voice benchmarks, taking #1 STT latency and #1 TTS word error rate in the mid-September run. (Nari Labs)

• HackerNoon explains Grok Voice Realtime, xAI’s audio-to-audio model with tools, search, and WebSocket streaming for voice agents. (HackerNoon)

• A tutorial builds a Vapi + ElevenLabs appointment agent that books slots and sends reminders, wired to Google Calendar and Twilio via n8n. (dev.to)

• A guide to fine-tuning NVIDIA Nemotron 3.5 ASR walks through adapting the open model to a language, domain, or accent with NeMo. (dev.to)

• A paper documents building a production Greek-English recognizer, evaluating 23 training runs against nine WER, LID, and hallucination gates. (arXiv)