Mitsubishi Electric unveils single-model audio separation tech

The announcement is part of Mitsubishi Electric's broader effort to position AI models closer to the point of data capture. The company says it developed the technology with Mitsubishi Electric Research Laboratories in the United States, and that the model can be input directly into a large-scale language model by encoding the separated sound.

The technology fits with Serendie, Mitsubishi Electric's digital platform, where linking separated audio to AI systems for anomaly detection or voice recognition could turn everyday sound into data for diagnosis and monitoring. Because it separates sound without training a separate model for each sound type, the company says it can run on edge devices and robots where computing power is limited, and it can work on various input devices without relying on specific microphones.

Whether the technology reaches those settings will depend on how well it performs when deployed through Serendie and at the physical AI stage targeted for around fiscal 2027. The practical test is likely to be whether linked anomaly detection and voice recognition systems can use the separated audio reliably enough in real environments.