Deeply to Present Microphone Acoustic Imbalance Resolution Technology at Interspeech 2025

Deeply, a leader in the domestic acoustic artificial intelligence (AI) sector, has proven its global technological prowess by announcing consecutive research achievements at major international conferences. Deeply unveiled its 'Unified Microphone Conversion' technology at 'Interspeech 2025', a speech conference held in Rotterdam, Netherlands, last August. This innovative technology solves the problem of acoustic data inconsistency arising from different microphones or recording devices, eliminating a major cause of decreased AI accuracy. Notably, by utilizing the Feature-wise Linear Modulation (FiLM) technique, it implemented 'Many-to-Many conversion' to process various microphone characteristics with a single AI model, improving recognition accuracy by 2.6 percentage points compared to existing methods and reducing performance discrepancies between devices by 0.8 percentage points. It achieved high performance using only synthetic data without actual data collection, drawing significant attention for demonstrating the potential to reduce commercialization costs and time. Earlier, at 'ICASSP 2025', a signal processing conference held in Greece in April, Deeply showcased its technological prowess with the acceptance of two papers. The first paper presented a technology based on the large-scale speech model 'Whisper' that detects help-request voices with a high accuracy of 88.5% even in noisy environments. A multi-task approach, which simultaneously learns speech recognition and noise classification, was applied. The second paper, 'ViolinDiff', is a sound synthesis technology that leverages pitch bend information from violin performances to achieve a realism score of 70.96 points, receiving favorable reviews in listening tests. Deeply integrates these accumulated advanced research achievements into its platform, 'Listen AI', applying them to various industrial sites, including domestic and international manufacturing lines, public safety networks, and smart city projects. Equipped with its own anechoic chamber, acoustic data collection system, and high-performance computing infrastructure, and through systematic R&D led by master's and Ph.D. level researchers, it performs the entire process from research to deployment in-house, enhancing technological completeness. Acoustic AI is now expanding beyond simple speech recognition into a wide range of areas such as machine state diagnosis, hazard detection, and improvement of industrial efficiency and safety, emerging as a core driver for future industries like predictive maintenance in manufacturing, urban safety, and content creation. A Deeply official stated, "In an era where AI utilizes sight and hearing, 'sound' is an important signal that alerts us to on-site dangers," adding that the company will actively pursue global collaborations and business expansion, leveraging these conference presentations as a springboard.