LG Uplus Corp. is ramping up efforts to improve its artificial intelligence (AI)-based voice recognition technology to enhance call quality and expand its AI agent services.

The company has built a 24/7 testing facility in Seoul dedicated to evaluating and advancing the performance of its AI call agent, known as ixi-O, including its ability to detect regional dialects and respond to unexpected call scenarios such as spam.

A recent visit to LG Uplus’s Quality Innovation Center in Magok, western Seoul, offered a glimpse into the company’s AI voice lab.

As the door to the test chamber opened, a phone rang and a synthetic voice responded, “How about grabbing coffee at six?” The conversation occurred amid background noise resembling a subway station or a bustling outdoor scene—but there were no humans present.

“An AI-based automation system runs throughout the entire testing process, from data measurement to diagnostics and evaluation,” said Kang Sung-kwang, an official from LG Uplus.

“We also simultaneously test three major third-party AI agents commonly used in the market to benchmark performance,” he added.

The test facility consists of two soundproof booths to simulate two-person phone conversations.

Each booth is equipped with four smartphones that pair up for calls, all handled entirely by AI—from dialing and answering to logging results. “The system automatically adjusts variables like background noise, voice levels, and the distance between smartphones, while the audio is continuously played back for testing,” said Moon Yoo-seok, a lead researcher. “The performance data is stored on servers and analyzed automatically.”

To simulate human voices, the lab uses a Head and Torso Simulator (HATS), allowing it to generate voice samples across gender, age, and regional dialects without hiring human models.

The lab is also fitted with ambient noise generators to replicate real-world call environments.

The system compares actual spoken phrases with what the AI transcribes, calculating accuracy on a syllable-by-syllable basis. Since July last year, the lab has tested roughly 8 million voice samples this way.

“The voice recognition accuracy has exceeded 99 percent,” Moon said. “We’ve also developed features that allow the AI to handle unpredictable situations, like hanging up on spam calls.”

This investment in quality stems from CEO Hong Bum-shik’s first message of the year for field employees, emphasizing a return to basics. “We’ve launched a task force to make call quality—fundamental to telecom services—our top priority,” Kang said.

[ⓒ 매일경제 & mk.co.kr, 무단 전재, 재배포 및 AI학습 이용 금지]