Hello, this is Sungmin Woo from ENERZAi. At ENERZAi, we buildon-device AI solutionsthat deliver strong performance even on hardwarewithout dedicated AI chips. Frommodel design and trainingtomodel compressionandcompile-level optimization, we provide thefull-stackpipeline required to implement accurate, fast, and efficient on-device AI.

ENERZAi works across a wide range of domains that need on-device AI such as automotive, wearables, drones, media/broadcast, and more. Among these, smart home is a key focus area. Recently, we deployed our on-device AI solution to two million TV boxes in Korea. Today, we’ll share how AI is being applied to smart home and which challenges ENERZAi’s on-device AI can solve.

TL;DR

• In smart homes, AI is used for many purposes such as energy optimization, security, and more. In particular, AI voice interfaces are being actively adopted thanks to their overwhelming convenience.

• Most AI voice interfaces today are cloud-based, which drives ongoing operating costs and entails network dependency, privacy concern, and latency. To address these issues, the industry is moving toward on-device AI.

• Realizing high-performance AI on resource-constrained hardware without a dedicated AI chip such as GPU/NPU is a challenging task.

• As a full-stack on-device AI company, ENERZAi provides high-performance, fast, and efficient AI for on-device environments by combining advanced quantization technologies with our proprietary AI inference optimization engine, Optimium.

• ENERZAi is delivering a wide range of on-device AI solutions for the smart-home market including voice control, live translation, voice agents, and more. If you’re exploring on-device AI for devices that seem “too constrained” for AI, ENERZAi would be a good partner!

AI@Home

“Turn on the living room lights.” “Turn off the TV.” “What’s on my schedule today?”

Homes that respond to voice are no longer a science fiction. They’re becoming everyday reality. Traditionally, “smart home” refers to connectivity and remote control built on IoT: operating appliances like air conditioners and washing machines via phone/tablet, or linking a home camera to your phone for remote monitoring. With Matter, the global interoperability standard for smart homes, the ecosystem is growing even faster.

Source: e4ds news

At smart home domain, AI is currently being used for a variety of purposes, such as energy optimization powered by usage pattern analysis and smart security powered by real-time video monitoring & anomaly detection.

Among these, AI voice interface is drawing particular attention. They enable hands-free control of lighting and appliances without manual input, and provide an intuitive UX that requires no user manual. As a result, they are spreading quickly across the smart-home domain.

Today, we would like to talk about the AI voice interface, which is located at the center of AI innovation in the smart home ecosystem.

Common Use Cases

The most common use case for voice interface in the smart home domain is voice-based command control. Beyond smart control panels in charge of household control or standard appliances (TVs, air conditioners, refrigerators), voice control is being applied to a wide range of devices including toilets, massage chairs, and care robots. Recently, more diverse AI capabilities such as live translation, search, and self-help assistant.

Smart TVs/TV Boxes

• Power/channel/volume control

• Activation of captions, voice guidance, and child protection mode

• Switching inputs and launching apps (ex. YouTube)

• Content search

Source: LG Electronics

Smart Control Panel

• Control of lighting, temperature, appliances, and security systems

• Search for weather and traffic information

• Booking community facilities

Source: Control4

Smart Refrigerators

• Adjusting settings (color modes, temperature)

• Automatic door open/close

• Food-related Q&A (ex. “What mode should I use to keep meat fresh for a week?”)

Source: LG Electronics

Smart Massage Chairs

• Start/stop massage

• Change modes and adjust intensity

Source: Brams

Why On-device AI?

Most voice AI functions in today’s smart homes are built on platforms like Alexa and Google Home. In these architectures, audio captured by the device is sent to the cloud for AI processing, and results are returned to the device.

Source: Amazon Developers

While the cloud facilitates AI adoption thanks to its abundant compute resources, cloud-based AI brings high operating costs (API calls, server instances), plus network dependency, privacy concerns, and latency from request/response cycles.

To overcome these limitations, we need powerful on-device AI technology to deliver high performance even on low-spec hardware with limited memory/compute resources.

Why On-device AI Matters

ENERZAi in Smart Home

Despite its clear advantages, implementing high-quality on-device AI voice interface is challenging since most home appliances use low-spec chips with limited memory and compute power. Research on model compression is active, but the goal isn’t just “smaller models.” We must achieve high-performance, robust AI on constrained hardware for positive user experience, which is challenging.

ENERZAi can tackle this challenge since we understand not only AI algorithms but also the hardware environments on which they run. Through extreme-low-bit quantization (Model compression technique that represents weights and activations with fewer bits to reduce memory and compute), we dramatically compress the AI model without compromising accuracy. Then, we conduct hardware-aware optimization using our proprietary AI inference engine, Optimium, to efficiently utilize the given hardware resources. This approach enables powerful AI models with exceptional speed, memory efficiency, and low power consumption, even on constrained hardware without dedicated AI chips including GPU or NPU.

Recently, we deployed our on-device voice control solution to two million LG Uplus TV boxes, delivering a significant reduction in operating costs compared to cloud-based voice control.

IPTV Providers Boost AI Integration for Enhanced Services

A key strength of this solution is that it achieves strong performance while using <100 MB of memory on the TV box SoC. Based on our unmatched competence and experience in AI optimization, we successfully met the client’s challenging requirements. Please refer to the post below for more details on our technology.

1.58-bit Quantization — the Wegovy for AI models 💉

ENERZAi aims to bring our on-device AI to more environments and more devices. As noted in earlier posts, we are developing a wide range of on-device AI models including Whisper and have implemented pipelines not only for voice-based command control, but also for live translation and self-help assistant.

AI is no longer confined to data centers or high-end PCs. Our goal is to make everyday devices (such as massage chairs, smart control panels, care robots, air conditioners, washing machines, and more) smarter and more convenient with our high-performance on-device AI solutions.

If you have questions about our technology and solutions, or if you’re considering on-device AI for devices that may seem “too constrained” for AI, please feel free to leave us a message anytime via the link below!

ENERZAi — Contact