Voice AI is moving into the TV, the meeting and the phone at once

Ordering from the TV

SoundHound AI and LG Electronics have announced a strategic partnership to expand commerce opportunities on LG Smart TVs in the U.S., the company says. The initial footprint includes approximately 20 million LG Smart TVs, with restaurants as the first category of focus and plans to extend to retail and other merchant categories over time. SoundHound provides transaction routing and secure checkout for the platform, and its agentic platform is named OASYS. m:…0444

The pitch to restaurants is low friction. The platform integrates directly with existing restaurant point-of-sale systems and kitchen execution systems, and participating restaurants require no new in-store hardware, the company says. SoundHound says it currently serves over 15,000 restaurant locations across the U.S. and serves brands including Jersey Mike's, Five Guys and White Castle. This suggests the TV is being treated as another ordering channel, not a new device category: the bet is that voice ordering works if the restaurant side changes as little as possible. m:…0444

Checking who is really on the call

As voice and video become transaction channels, impersonation follows. Resemble AI has launched a native real-time deepfake detection integration for Microsoft Teams, the company says, which runs as a Teams app without requiring a separate bot to join the meeting. It includes four detection models, named Detect, Identity, Signal and Intelligence, and it checks that each person on a call matches the Microsoft 365 account they signed in with. Deployment happens through the Teams admin center. m:…0144

The cases Resemble cites explain the urgency. In 2024, an Arup employee authorized 15 transfers worth $25.6 million after joining a video call where his CFO and colleagues were all deepfakes. In 2025, a Singapore finance director transferred more than $499,000 after a deepfake executive call, and in February 2026 a cloned voice helped move roughly 95 million euros out of Fideuram, part of Intesa Sanpaolo. Gartner's September 2026 survey of 297 CISOs, as cited by Resemble, found 41% had dealt with a deepfake used in a social-engineering attack on an employee phone call in the previous year and 36% on a video call. Resemble says its audio model ranked first on the Podonos leaderboard in August 2026 with 99.47% accuracy, a 0.4% miss rate and a 0.7% false-alarm rate, and that customer data is not used to train its models. m:…0144

Siri opens to developers

Apple has released a Developer Preview for Siri AI integration starting with the 27.3 beta versions of iOS, iPadOS, macOS and visionOS, according to Apple Developer News. Developers can use App Intents to bring app content and actions into Siri AI, and they can teach Siri AI to combine multiple actions into a single experience using new skills. Siri AI can surface any entity in the Spotlight semantic index, and it supports personal context understanding, onscreen awareness and natural language processing. m:…5593

The limits are stated plainly. Siri AI is currently unavailable in the EU for iOS, iPadOS and watchOS, it is not available for users under 13, and server-side Apple Intelligence features are subject to daily usage limits. Apple plans to expand language support to include Traditional Chinese, Norwegian, Turkish and Vietnamese, and App Store submission for Siri AI integrations opens in the spring. For creators and developers, the practical question is whether voice becomes a distribution channel for apps, or stays a way to reach content already inside them. m:…5593

Voice input reaches Linux

Superwhisper has launched a native voice input application for Linux, the company says. It supports local processing on CPU or GPU, as well as cloud processing for maximum speed, and it inserts text directly into the focused application without window switching. It is configured through a graphical interface, terminal UI or CLI commands, and it supports Omarchy, Arch, CachyOS, Fedora, Ubuntu, Debian and openSUSE. It currently requires Wayland. m:…8288

The company frames the gap it fills in plain terms: Linux voice input methods previously required stitching together recording tools, shell shortcuts and clipboard hacks. Taken with the TV, meeting and phone moves above, this suggests voice is consolidating into a normal input method across every screen, with local processing offered as the privacy answer and detection offered as the fraud answer. The open question for all four moves is the same: whether users actually want to talk to each of these surfaces, or whether voice stays a niche input for specific tasks. m:…8288