Example of TUSS-based system

TOKYO, August 31, 2026 – Mitsubishi Electric Corporation (TOKYO: 6503) announced today the development of its Task-Aware Unified Source Separation (TUSS) technology, which enables a single AI model to separate and extract desired sounds from complex acoustic mixtures. Integrating the technology into physical AI systems in acoustically complex environments, such as manufacturing sites and public spaces, helps improve the reliability of these systems by enabling them to understand complex situations more accurately. Mitsubishi Electric developed the TUSS technology jointly with Mitsubishi Electric Research Laboratories, Inc. in Cambridge, Massachusetts, USA.

In various environments, tasks rely on sound for purposes such as voice-controlled equipment operation, anomaly detection and on-site situational awareness. However, these tasks can become difficult when target sounds from humans, machinery, etc. are masked by other sounds.

The TUSS technology uses prompts to specify the type and number of sound sources to be separated or extracted, enabling a single AI model to perform various sound-source separation tasks, including speech separation, speech enhancement and environmental sound extraction. Unifying these tasks in a single model avoids the need to develop separate source separation models for different target sounds or applications, allowing the technology to adapt flexibly to varying acoustic environments and operational requirements. Linking the extracted sounds to AI applications, such as anomaly detection, speech recognition, voice-controlled equipment operation and operational recordkeeping, enables these systems to understand complex situations more accurately.

The technology will be exhibited under the title “Ear of Physical AI: Distinguishing Diverse Sounds, from Machine Anomalies to Human Voices” during CEATEC 2026 at Makuhari Messe, Japan from October 13 to 16. A live demonstration will process mixed audio captured at the venue, including sounds from machinery, musical instruments and speech from multiple speakers, and the technology will separate and extract prompt-specified individual sound sources for speech recognition and abnormal sound diagnosis.