Toshiba develops AI technologies for acoustic and speech recognition

Toshiba has announced the development of two new AI technologies designed to support the realization of Physical AI by accurately recognizing equipment sounds and worker speech in industrial environments.

The first technology, termed ‘Few-shot acoustic event detection technology,’ allows users to detect specific sounds, such as equipment malfunctions or alarms, by registering only a few samples. This eliminates the need for large-scale data collection or continuous model retraining, making it adaptable to different sites and noisy environments. In benchmark evaluations, it achieved top-tier performance without additional training.

The second technology, ‘Error correction record-utilizing speech recognition technology,’ improves the accuracy of recognizing specialized terminology and proper nouns. By utilizing records of corrections made by users and integrating them with Large Language Models (LLMs), the system can correct misrecognitions in real-time without requiring full model retraining or manual dictionary updates.

Toshiba intends to apply these technologies across various sectors, including manufacturing, energy, social infrastructure, logistics, and construction, to support digital transformation (DX) and the transfer of expertise from veteran workers to new employees.