Top Updates đȘ
A report put the audio AI market near $117B by 2033, a 19.3% CAGR led by generative voice platforms. GlobeNewswire
ElevenLabs released Scribe v3 Live, a realtime STT model that uses conversation context to avoid cutting callers off. The Decoder
Boson AI previewed Higgs Realtime Duplex, a 45B model that listens, speaks and renders a face at once. audioXpress
Bland introduced Tofu-1, a full-duplex speech-to-speech model aimed at high-stakes banking and healthcare calls. MarkTechPost
DeepL unveiled Aure, a voice-translation model that carries a speakerâs tone, emotion and intonation across languages. The Decoder
Retell launched Tell-1, an in-house LLM for live contact-center calls that reasons while it speaks. Smart Customer Service
Zyphra open-sourced ZONOS3-Mini, a 100M-active-parameter TTS model that clones voices in 43 languages on-device. MarkTechPost
ThunderPhone launched v3 and Zeus, running several STT and LLM models per call with an agent that tunes them. PR Newswire
Retell hit a $100M run rate, per the company, with over 100M calls a month. Forbes
PromptPath raised $12M to coach dealership staff in real time rather than replace them with voice AI. citybiz
Krispâs CEO expects India to become its largest deployment market, led by voice translation in contact centers. Business Standard
Bandwidth previewed Call Intelligence and its own STT model, with GPUs deployed inside its phone network. Dealroom
Google brought Gemini voice to Apple CarPlay, with hands-free Gemini Live conversations from the car screen. Bloomberg
Cresta launched Journey Intelligence, linking calls and chats into one continuous customer journey for agents. PR Newswire
Quiqâs voice AI landed on Zoom Contact Center, available through the Zoom App Marketplace. Smart Customer Service
XTIUM launched XTAI Voice, a managed AI voice service that answers calls and escalates to humans. Telecom Reseller
Voiceware introduced iGentrix, an AI agent builder pitched to MSPs and resellers for vertical customization. Telecom Reseller
Modulate upgraded Velma with synthetic-caller detection, backed by a new Twilio Ventures investment. Unite.AI
Neural Defend tied for first on a deepfake-audio benchmark, topping Podonosâs independent leaderboard. Business Standard
Kaspersky flagged a Muse AI zero-day, a flaw that let attackers hijack the assistantâs voice server. Kaspersky
Antimattr is building noise-proof voice wearables, a ring-and-earphones combo for talking to AI agents. NDTV Profit
CNET tested Apple Watch Audio Intelligence, a hands-on look at the privacy tradeoffs of always-on listening. CNET
Gartner named speech reasoning models a top 2027 trend, projecting 70% of enterprise voice calls autonomous by 2030. Gartner
Engineering Corner đ
Google added on-device STT and TTS to AI Edge Gallery, bundling Parakeet, Kokoro and Phonon for offline voice. Google
Firefox 160 added on-device speech recognition, real-time transcription in 40 languages via the Web Speech API. Phoronix
Cursor and Antigravity gave coding agents a voice, two-way calls on Grok Voice and Gemini Live. The New Stack
Liquid AI detailed LiquidCodec, a 74M streaming audio codec encoding speech ~361x faster than real time on CPU. MarkTechPost
A paper introduced Pandemonium, long-form target-speaker ASR with cross-speaker attention for meetings. arXiv
Diffio shipped a speech-restoration model, diffusion-based cleanup for noisy interviews and podcasts. Radio World
A walkthrough covered deepfake voice-clone defense, a municipal-fraud scenario for voice agents. dev.to
A guide argued a familiar voice is not authorization, 12 tests for voice agents in group calls. dev.to

