ElevenLabs V4 and Microsoft V2 top the charts

Voice AI weekly digest

Krisp publishes an open Voice Isolation benchmark, testing 11 STT configurations on real audio with a second voice, where word error rate drops 73%.

Events 🎤

• Vonage and Deepgram open Deepgram’s SF HQ for “Voice AI in the Wild” (Oct 6, San Francisco, Partiful)

• Agora runs “From Prototype to Production: Scaling Voice AI” (Oct 6, San Francisco, Partiful)

• AssemblyAI hosts a hardware hackathon (Oct 7, San Francisco, Luma)

• Aqua Voice, Speechmatics, and LiveKit tackle “Solving voice as an interface” (Oct 7, San Francisco, Partiful)

• Coval and Cartesia host “Voices in the Room” (Oct 8, San Francisco, Partiful)

• Agora brings EdTech leaders together on voice-first learning (Oct 8, London, Luma)

Top Updates 💪

• ElevenLabs launches Eleven v4 and v4 Turbo, adding automatic expression control, 10-second voice cloning, and 90+ languages (TechCrunch)

• Microsoft ships MAI-Transcribe-2-Streaming plus MAI-Voice 2.1 models, topping the streaming AA-WER benchmark at 2.5% (Unite.AI)

• Alibaba releases Qwen-Audio 3.1 Realtime, a full-duplex voice model with 262K context and tool calling, cutting voice API prices up to ~85%. (MarkTechPost)

• Inworld acquires Ultravox (formerly Fixie) to add a real-time voice-agent platform with turn-taking and interruption handling to its speech stack. (Morningstar)

• Wemnal integrates Krisp VIVA to improve its voice AI agents (LinkedIn)

• Modulate raises $25M for voice models and an analysis suite spanning transcription, emotion, deepfake and AI-music detection, and agent policy enforcement. (TechCrunch)

• Instinct raises $1B at a $10B valuation for a personal AI agent that calls and texts on users’ behalf, quadrupling its valuation in a month. (SiliconAngle)

• Relay raises $36M to turn the frontline push-to-talk radio into an AI assistant that logs and acts on what workers say. (PYMNTS)

• Presto raises $10M to expand its drive-thru voice AI across enterprise QSRs, now live in hundreds of locations. (Yahoo Finance)

• Deepslate raises $8.8M to build a Europe-hosted speech-to-speech model it bills as the fastest on Artificial Analysis at 440ms. (Tech Funding News)

• Klang raises €1.32M and open-sources Pianissimo, a Swedish speech-to-text model that runs locally and builds on NVIDIA Parakeet. (EU-Startups)

• ElevenLabs doubles its valuation to $22B, letting employees cash out via a $300M tender offer six months after its $11B raise. (TechCrunch)

• AI could bring 10x more Voice work to India (BusinessWorld)

• Fireflies launches Fireflies Talk, free unlimited voice dictation across apps on Mac and Windows in 100+ languages. (GlobeNewswire)

• Tavus previews Griffin, a full-duplex video-to-video model that sees, hears, and talks back, with 48% of testers thinking it was human. (The Rundown AI)

• Suno launches Speech beta, generating synthetic voices with matched background music in one track from a prompt or script. (The Verge)

• Mobvoi announces the TicNote Watch, a $249 smartwatch that records, transcribes, and summarizes conversations from the wrist offline. (PR Newswire)

• EssilorLuxottica launches Nuance Audio Plus, second-gen open-ear hearing glasses adding iPhone calls and longer battery, from $829. (Wareable)

• Starkey unveils Omega AI+, its most advanced hearing aid, built on a new Gen AI neuro processor and quad-DNN architecture. (Fortune)

• Logitech launches the Zone Vibe Pro, a boomless headset with AI call noise reduction filtering up to 93% of background sound. (CravingTech)

• RingCentral partners with NovelVox to wire contact centers into core banking platforms like Fiserv, FIS, and Jack Henry. (RingCentral)

• An AI voice test could screen for Type 2 diabetes, flagging risk from 20 seconds of speech in a 21,000-person study. (Forbes)

• Deepfake defense draws new capital, with voice-detection checks forecast to near $5.5B by 2028 as fraud screening scales. (Biometric Update)

• Telecom Reseller argues contact centers should own their voice layer, with Gladia’s CEO framing voice AI as core infrastructure, not an add-on. (Telecom Reseller)

---

Engineering Corner 😎

• FermionResearch releases Phonon-2, the most accurate open English ASR model under 900MB at 5.21% WER, in a 164MB download. (Hugging Face)

• A comparison of OpenAI, Gemini, and Qwen realtime voice APIs finds up to a 213x gap in per-minute cost across the three. (Tech Insider)

• whisper-local ships offline push-to-talk dictation, a free faster-whisper tool that types at the cursor in any app, fully on-device. (PyPI)

• A paper proposes inference-time target-speaker unlearning in ASR, letting a speaker opt out of meeting transcription without leaving the call. (arXiv)

• Linux Journal shows adding speech synthesis to scripts and systems, using TTS APIs to voice Linux services without hosting a model. (Linux Journal)

• A breakdown of voice-agent latency explains why response delay kills calls and how to cut it through the STT, LLM, and TTS stack. (AI in Plain English)

• A dev essay on why AI voices wear thin after three minutes dissects the “30-second trap” of long-form synthetic narration. (dev.to)

• Perficient briefs on NotebookLM and Wispr Flow, pairing source-grounded research with voice dictation for content workflows. (Perficient)