Deeply’s Listen AI Achieves 99% Accuracy and Targets the Manufacturing AI Market

Deeply CEO Suji Lee recently told Maeil Business Newspaper that manufacturing-focused AI could become one of Korea’s strengths in the AI era. Unlike general-purpose AI that excels at conversation like a chatbot, she said there is a need for AI that can actually be applied on manufacturing production lines. Deeply is targeting the manufacturing AI market with its “Listen AI” solution, which analyzes industrial sounds to inspect quality. The area Deeply is focusing on is the inspection process that previously depended on human ears. In manufacturing sites, workers used to judge defects by listening directly to machine sounds or by placing a stethoscope-like tool against the equipment. Deeply is transforming this hearing-based inspection into AI-based data analysis.

The key is learning normal data. Because defective data is difficult to collect in sufficient volume, Deeply first trains on large amounts of sound from normal products and then detects abnormal sounds that fall outside the normal range. Lee explained, “We mainly use anomaly detection methods that define the normal range first and then identify anything outside that range as abnormal.”

Deeply’s strategy is to build its own AI models. Industrial sound data does not have as many public general-purpose models as speech AI does. In addition, models that perform well in laboratories or quiet environments often lose accuracy in real production settings because of equipment vibration, background noise, and variations from line to line. Lee said, “Open-source models may perform well in limited environments, but accuracy often drops significantly when they are introduced to mass-production lines. We determined that our own model, which can be continuously adapted to each site, was essential.”

Based on this field experience, Deeply is also improving a common acoustic foundation model. The company is working to evolve frequently encountered sound types, such as rotating machinery sounds and the “click” sounds generated during fastening and assembly processes, into a shared sound foundation model. As a result, deployment that once took several months on-site can now be introduced within about a week for the basic product version.

This year, Deeply has shifted its business focus toward quality inspection and defect detection on manufacturing lines. The company says that after its solution is introduced at one line, it is increasingly expanded to additional lines. Lee said, “Once our accuracy began reaching the high 99% range last year, customers started making repeat purchases and expanding the solution to other lines. It is rewarding to see that we are being recognized in the field as a proven product.”

Deeply is also seeing more contact with global manufacturers, not just domestic ones. The company says that manufacturing AI use cases created first in Korea are now being reported up to global headquarters. Lee said, “In the past, the structure was usually that headquarters’ methods were simply passed down to local subsidiaries. Recently, however, Korean manufacturing AI use cases are being reported back up to global headquarters. There is now a growing perception that Korea is strong in manufacturing AI.”

Deeply plans to make next year its starting point for overseas expansion. Rather than only following domestic manufacturers when they build overseas plants, the company is now also seeing cases where foreign companies are directly considering contracts. Deeply is therefore discussing agreements with overseas partners in earnest. Lee said, “Deeply wants to become a company that turns unstructured data such as sound into manufacturing competitiveness.”