On-device voice artificial intelligence (AI) startup PrelisKuner announced on August 20 that its paper was selected for an oral presentation at the world-renowned voice AI conference 'Interspeech 2025', based on its outstanding technological capabilities.
Interspeech is the largest academic conference in the field of voice AI, with thousands of voice AI experts from around the world, including major big tech companies such as Google, Meta, and Apple. This year, the event was held in Rotterdam, the Netherlands.
PrelisKuner's paper covers a new 'Parameter-efficient Fine-tuning' technique that enables easy conversion of streaming voice recognition models into highly accurate non-streaming models. This technology allows 'real-time voice recognition models' to be quickly and cost-effectively transformed into 'offline analysis models that prioritize accuracy.'
While achieving a level of accuracy similar to existing non-streaming models, the additional parameters required remain below 10%. After training just one streaming model, it can be instantly utilized as a non-streaming model as needed without additional large-scale training, thereby increasing development and operational efficiency.
PrelisKuner is developing 'GrassLand,' an engine composed of an on-device voice AI foundation model that can run directly on commercial devices, along with optimization technologies that can be easily applied to various domains.
Kim Kiyeon, CEO of PrelisKuner, stated, "This research presents an essential technology for the on-device AI market, which requires high performance even in resource-constrained environments," adding, "We will continue to advance our proprietary engine 'GrassLand' and successfully apply our technology across various industries."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
