SoundHound ships OASYS Edge for on-device agentic voice AI

SoundHound AI introduced OASYS Edge, a new embedded architecture that brings LLM-powered voice AI agents onto edge hardware for vehicles and smart devices. Agents can keep talking locally when the network drops, keep cabin audio on-device by default, and hand off to the cloud only when needed.

SOFTWARE desk — a voice agent in a car or a gadget still goes quiet in a garage, a tunnel, or a stretch with no network, and every utterance sent to the cloud is a privacy choice plus a bill the manufacturer pays. SoundHound’s announcement puts that agent on the device, so the conversation can continue offline, cabin audio can stay there, and the cloud is used only when the task needs it.

What the lede says the launch is. SoundHound AI, called a global leader in voice and agentic AI, today announced OASYS Edge, a new embedded architecture that brings LLM-powered voice AI agents directly to edge hardware. LLM means a large language model, the kind of model that can follow a spoken request in ordinary words. Edge, here, means the car or the gadget itself, not a data center. An agent, here, is software that can take the next step on a task, not only answer one question. Those glosses are this desk’s. The page prints LLM and does not expand it. Available for vehicles and smart devices, OASYS Edge provides proactive conversational AI assistance that runs regardless of connectivity. Because processing happens locally, conversations stay private by default, and can continue uninterrupted even when network signals are unavailable. Global leader, proactive, and private by default are the release’s words. This desk did not drop a phone into a tunnel.

The first-time sentence, and whose survey that is not. OASYS Edge marks the first time global automakers, device manufacturers, and service providers will be able to deploy fully embedded agentic AI, with SoundHound setting the global pace as a longtime pioneer in edge and hybrid processing. First time, global pace, and longtime pioneer are the release’s claims. They are not a product census this desk made. Agentic AI, here, means the software can decide the next action, not only match one spoken command to one button. Hybrid, here, means some of the work stays on the device and some goes to the cloud. Those glosses are this desk’s.

Build once, and what the release says that control is for. With OASYS Edge, developers from any business can build intelligent voice agents on the OASYS platform once, then deploy them in the cloud, on the edge, or in a hybrid environment. The release says this flexible architecture gives manufacturers full control over their technology stack for stronger data privacy, lower cloud costs and bandwidth needs, and consistently high-speed responsiveness. Full control, and consistently high-speed, are the release’s words. They are not a bill or a stopwatch this desk ran. For drivers and end users, the release says OASYS Edge means a fast, reliable voice experience that feels unlimited, with AI agents that can understand even vague requests, resolve complex needs, and make smart decisions anytime in any location. Feels unlimited is the release’s phrase. It is not a test this desk drove.

James Hom, as a quote, not as a measured saving. James Hom, Chief Product Officer and Co-Founder at SoundHound AI, said that for a long time the industry assumed this kind of sophisticated agentic capability could only live in the cloud, forcing companies to compromise on speed, cost, reliability, and privacy. He said OASYS Edge fundamentally shifts that dynamic. He said that by enabling developers to create AI voice agents just once and seamlessly deploy them in the cloud, on the edge, or across hybrid environments, they give manufacturers unprecedented flexibility. He said this allows them to provide uninterrupted AI assistance to end users without network overhead or cloud processing. Unprecedented, and without network overhead, are his sentences. They are not a cost study this desk signed. The page this desk read ends that speech there. This desk did not interview him.

What the release says happens on the device, and the two examples it prints. OASYS Edge goes beyond rigid, single-action voice commands to deliver natural, multi-system orchestration locally, and only hands off tasks when cloud connectivity is needed. Orchestration, here, means the assistant coordinates more than one system from one request. That gloss is this desk’s. The first example is a driver heading through a remote stretch with poor reception who says, “Navigate to Los Angeles and add two iconic tourist attraction stops along the way, and tell me about each of them.” Even with zero connectivity, the release says, OASYS Edge calculates the route, identifies well-known landmarks from its local knowledge base, and narrates highlights for each stop, updating the plan in real time, entirely on-device, with no cloud required. A knowledge base, here, is reference material stored on the device. That gloss is this desk’s. The second example is a home or workplace robot that needs to listen continuously. OASYS Edge handles query detection and everyday conversational responses entirely on-device, so the robot can stay ready without streaming audio to the cloud. Those two scenes are the release’s illustrations. This desk did not drive to Los Angeles, and it did not put a robot in a kitchen.

The five benefits the release lists for manufacturers and the people using the devices. An OEM, the heading’s word, is an original equipment manufacturer, the company that builds the car or the gadget. That gloss is this desk’s. Flexible hybrid architecture: developers can create AI agents once on OASYS and deploy across cloud, edge, or hybrid environments depending on performance, cost, and hardware needs. Instant responses and offline reliability: voice requests execute at fast, natural conversational speeds, and systems operate reliably in connectivity dead zones, such as underground garages, tunnels, and remote areas. Privacy by design: user interactions, voice processing, and telemetry can optionally remain on the local device. Telemetry, here, is the measurements the device records about the session. That gloss is this desk’s. The same bullet says no sensitive cabin or personal data is transmitted to external servers, providing full data sovereignty and meeting strict international privacy mandates. Optionally remain, and no sensitive cabin data is transmitted, are both on that bullet. Do not drop the word optionally. Sovereignty, here, is the release’s word for keeping that data under the manufacturer’s control. That gloss is this desk’s. Full, and strict, are the release’s adjectives. This desk did not read a privacy statute. Onboard multi-system orchestration: OASYS Edge understands intent, resolves ambiguous requests, and autonomously controls multiple features simultaneously without cloud reliance, powered by an optimized, purpose-built range of language models. The release does not name those models. Do not invent them. Lower operating overhead: bypassing constant cloud processing significantly reduces recurring API costs and bandwidth demands for manufacturers. An API cost, here, is the fee for each request sent to a cloud service. That gloss is this desk’s. With customizable model sizes and configurations, the release says the solution delivers high autonomy even on low-cost chipsets. Significantly reduces, and high autonomy, are the release’s claims. This desk did not price a chipset or an API call.

When the release says it ships. SoundHound AI’s OASYS Edge is scheduled for deployment in late 2026, with live demonstrations available now and to be featured on the show floor at CES 2027. CES is the consumer electronics show. That gloss is this desk’s. Late 2026 is a window, not a day. Demonstrations available now, on a page dated 24 Sep 2026, is the release’s tense. This desk did not attend a demonstration, and it did not book a CES booth.

Who the about box says the company is, and whose counts those are. SoundHound AI (Nasdaq: SOUN) is a voice and agentic AI company that enables businesses to deliver natural, end-to-end conversational experiences across phones, kiosks, chat, smart devices, drive-thrus, TVs, in-vehicle, and more. Its agentic platform, OASYS, is described as a self-learning, orchestrated AI system where organizations can build and deploy conversational AI agents to handle transactions, tasks, and workflows on behalf of customers and employees. Self-learning is the about box’s word. It is not a training log this desk opened. Built on proprietary technology backed by 750+ patents and years of AI research, SoundHound serves leading brands across automotive, financial services, healthcare, retail, telecommunications, and more. It powers millions of products and processes billions of interactions annually for enterprise customers worldwide. 750-plus is more than 750 patents. Millions and billions are the about box’s scale words. Those translations are this desk’s. The about box does not print the patent list or the customer names behind those totals. The learn-more line is www.soundhound.com. The page does not print a press email. This desk did not count the patents or the interactions.

What the card shows, and what the 24 Sep release does not print as labels. The card is the OASYS Manage Agents product screen on the newsroom page, the image whose file name is straightened-extended 1.avif. The still shows the OASYS wordmark, the line Orchestrated Agent System, and a Manage Agents grid. Cards on that grid are marked Deployed. Those words are the product UI’s. The release’s sentences about Los Angeles, the robot, late 2026, and CES 2027 are not printed as labels on this card. The card has no desk date on it. The schema image is a separate deep-plum press thumbnail. This filing does not use that thumbnail.

What the learn-more page adds, and what it does not let you file as the newsroom release. The newsroom’s learn-more link goes to the company article “OASYS Edge: embedding agentic intelligence into vehicles and smart devices,” dated September 24, 2026, with a 5 minutes label, by Pedram Faghfouri, VP of Product and Technology, Automotive and IoT. Five minutes is time to read. It is not a second publish clock. The article does not print James Hom, Nasdaq, 750 patents, late 2026, or CES 2027. It says businesses have had to compromise: rule-based embedded systems answer fast but match predefined commands, and cloud AI can reason through a complex request but depends on a connection and on operating costs that are hard to predict. It says agentic intelligence runs fully on board, and the assistant turns to the cloud only when a request needs live online information or a connected service. A vehicle example on that page says “It’s freezing in here” might be an open window or a heated seat that is still off, so the right response depends on context. Another example is a night drive in rain: one sentence, and local agents for wipers, windows, headlights, mirrors, media volume, and the dashboard are supposed to act together. A third example books a table: search locally, then use the cloud only for the reservation. Those scenes are the article’s. The newsroom’s scenes are the Los Angeles stops and the robot that does not stream audio. Do not collapse the two sets. The article says on-device requests do not carry the same per-request cloud cost, and that the language model can be tuned for the brand and the device. Those lines are the article’s. The newsroom’s cost line is lower recurring API costs and bandwidth. This desk did not tune a model.

What the product pages add, and which older lines they do not let you file as today. The OASYS page at soundhound.com/oasys, read 25 Sep 2026, calls OASYS the Orchestrated Agent System for conversational AI. It says SoundHound was named a Leader in the 2026 Gartner Magic Quadrant for Conversational AI Platforms, and the overall leader in the 2026 ISG Buyers Guide for Conversational AI Emerging Providers. Those two lines are that page’s. The 24 Sep newsroom release does not print them. The page this desk read does not print OASYS Edge, Santa Clara, late 2026, or CES 2027. The Edge and Cloud page at soundhound.com/oasys/edge-and-cloud, read the same day, does not print OASYS Edge either. It quotes Alexandre Fromion, Senior Expert, Connectivity and Infotainment, Stellantis, on hybrid cloud and embedded connectivity for always-on voice. It prints 1.7 beside an M, labeled vehicles with generative AI assistant. 1.7 million is this desk’s expansion of that M. The figure is that page’s. The 24 Sep release does not print Stellantis or 1.7 million. This desk did not audit the vehicle count, and it did not read the Gartner report.

Plain English for the rest of the card. A voice agent is software that listens and then does the next step. OASYS Edge is SoundHound’s name for running that agent on the car or the gadget, not only in the cloud. Build once means a developer makes the agent on OASYS and can then place it in the cloud, on the device, or in a mix of the two. When the network is gone, the release says the conversation can continue on the device. The privacy bullet says voice and telemetry can stay on the device, and that sensitive cabin data is not sent out. The word on that bullet is optionally. Garages, tunnels, and remote roads are the release’s examples of dead zones. The Los Angeles drive with two landmark stops, and the robot that answers without streaming audio, are the release’s examples. Late 2026 is the deployment window. Demonstrations are described as available now. CES 2027 is the show floor the release names. James Hom is chief product officer and co-founder. More than 750 patents, millions of products, and billions of interactions are the about box’s figures. 7:00 a.m. UTC is 3:00 a.m. Eastern. The Manage Agents grid is the product screen. This filing is the 24 Sep announcement.

PRIMARY here: SoundHound AI’s 24 Sep 2026 newsroom announcement, visible line September 24, 2026, schema datePublished 2026-09-24T07:00:00.000Z, datelined Santa Clara — Tier A PRIMARY, the company’s own page. STATUS PRIMARY. The OASYS Edge name, vehicles and smart devices, local processing, private by default, the first-time sentence, build once across cloud, edge, or hybrid, the Hom quote, the Los Angeles landmark example, the robot that does not stream audio, the five benefit bullets including optionally on-device telemetry and no sensitive cabin data transmitted, garages, tunnels, and remote areas, purpose-built models with no names printed, lower recurring API and bandwidth costs, low-cost chipsets, late 2026, demonstrations available now, CES 2027, the about box including 750+ patents, millions of products, and billions of interactions are that page’s. The rain scene, the freezing-cabin scene, the restaurant booking, Pedram Faghfouri, and the five-minute label are the learn-more article’s. The Gartner and ISG leader lines are the OASYS product page’s. The Stellantis quote and the 1.7 million vehicle line are the Edge and Cloud page’s. The Manage Agents grid and the Deployed marks are the product screen’s. NOT claimed: that this desk drove the Los Angeles route, that first time was surveyed, that optionally was dropped, that a model name was printed, a patent count this desk made, a dollar API price, a chipset price, a CES booth this desk booked, that 1.7 million or the analyst leader lines are sentences on the 24 Sep release, a login, a stock tip, or investment advice. The card is the Manage Agents screen. Distinct from the already-filed docker-cloud-sandboxes, kontext-agent-runtime, dataiku-agent-management, and gurucul-ai-risk-response.