최근 가전과 IoT 제품 등 일상 주변의 디바이스 등에 AI 기능이 탑재되고 있는 추세이다. 특히 이러한 엣지 디바이스에서의 AI 시스템은 엣지단에서 AI를 수행하는 칩을 둠으로써 메인 모듈 혹은 프로세서가 상시 켜져 있을 필요성이 줄어들어 전력 효율을 도모할 수 있다.
▲Zoom interview scene with Park Se-jin, CEO of Gamba Labs
Ultra-lightweight voice recognition and speaker recognition on-device AI
Heterogeneous MCU, VIOLA framework support construction
AI modeling must overcome misrecognition and noise environment challenges
Recently, AI functions are being installed in everyday devices such as home appliances and IoT products. In particular, AI systems in these edge devices can improve efficiency in power and product life by placing chips that perform AI at the edge, thereby reducing the need for the main module or processor to be always on.
Gambalabs is an on-device AI company with ultra-lightweight speech recognition and speaker recognition technologies, and has model lightweighting technology that enables porting AI models to edge chips such as MCUs. This lightweight AI model is an essential technology for on-device AI, and requires a high level of technical skills and know-how to reduce model capacity to kilobytes while maintaining relatively the same performance.
The latest trend is that major home appliance manufacturers are preparing to install keyword voice recognition in their next-generation home appliances and are reportedly looking for AI technology solution providers.
“Gamba Labs’ approach is to prioritize the high performance and efficiency of AI models, but also focus most on price competitiveness,” said Park Se-jin, CEO of Gamba Labs, who we met through an online Zoom interview. He also revealed that the company is currently developing its own hardware to optimize the hardware.
Park Se-jin, the CEO who aims to supply inexpensive on-device AI chips with the slogan of “$1 processor, $2 module,” cheerfully called it the “Daiso strategy,” but it was clear that it was a serious vision.
Currently, Gamba Labs has completed product testing of MCU companies such as Espressif, Renesas, and Nuvoton, and has established the VIOLA (Voice Interface Over Lightweight AI) framework that enables AI models to be automatically loaded onto heterogeneous MCUs.
There are various MCU manufacturers, and there are more than tens of MCUs released by a single manufacturer. In order to load AI models on these various MCUs, an AutoML framework that automates this is essential.
GambaLabs can quickly and efficiently perform AI model conversion optimized for MCU hardware of various specifications through the VIOLA framework. Connecting this framework to business would make us a SaaS solution provider that provides lightweight AI models, but this is not the business direction that Gamba Labs is pursuing.
▲Gamba Labs Demo Case / (Capture: Gamba Labs Homepage)
Gamba Labs is preparing to conquer the market by optimizing the voice recognition AI model it has developed to fit the hardware specifications and MCU desired by customers, and supplying chips/modules equipped with such models or the AI model itself to customers.
△It can be applied to various applications such as smart lighting, kiosks, robot vacuum cleaners, home and kitchen appliances, door locks, laptops, and automobiles, and the development of related products is progressing rapidly in the current market.
CEO Park Se-jin pointed out that “global MCU companies are providing AI modeling libraries and development frameworks, but they are merely providing a development environment,” and that “when you actually test them, you can’t be satisfied with the performance because they don’t take into account misrecognition and noise environments.”
Currently, the major issues in voice and speaker recognition are high misrecognition and response to noisy environments, and if there is no development know-how for these issues, the market competitiveness of the product is bound to decline.
CEO Park said, “The AI model size that goes on the MCU must be small, and we have reduced the voice recognition model to 30kbytes, and the appropriate size is around 50~60kbytes.” He added, “The mass-produced module optimized for voice recognition has been developed, and the one for speaker recognition has been developed.” “The module is in the final stages of development, and ultimately, we are also developing an ultra-small artificial intelligence processor that can recognize voices and speakers simultaneously,” he said, expressing his confidence.
As small and medium-sized manufacturers and large corporations are developing on-device AI products, an ecosystem of companies with lightweight AI solutions is blossoming. As the activities of startups in the on-device AI ecosystem are expected to increase, it will be interesting to see which AI startups will emerge as the “spittle in the bag.”