AI RESEARCH
휴먼교육센터와 휴먼딥브릿지가 연구·출간하는 인공지능 논문을 소개합니다.
Vision transformer-based emergency classification in Azure cloud
In modern society, due to the rapid aging of the population, health and long-term care (LTC) management of the elderly aged 65 and over are important social issues. Existing machine learning models that analyze standardized features have limitations in effectively interpreting complex bio-signal patterns. To solve this problem, we propose a vision transformer-based emergency classification service. A new foundation model based on the Transformer architecture can understand human language more accurately and classify requests with much less training than before. We modified the existing Transformer model MobileHART to be suitable for emergency classification.
A Corpus-Based Approach to Candidate Generation for Korean Dialect ASR Post-Processing
Large-scale Automatic Speech Recognition (ASR) models demon- strate high performance on clean standard Korean speech, but transcription errors frequently occur in Korean speech that includes regional dialects, colloquial re- ductions, and contracted forms. Generative LLM-based post-processing can sig- nificantly improve automatic metrics, but it carries the risk of introducing expres- sions absent from the source text or excessively altering sentence structure. This study proposes a controlled post-processing pipeline that combines explicit knowledge-based candidate generation and candidate selection for Whisper- based Korean dialect ASR transcriptions. The proposed method generates up to five correction candidates using training error pairs, a dialect dictionary, exam- ples of single phonological changes from a pronunciation rule corpus, and self- candidates, and selects the final output exclusively from within the candidate list. In held-out test utterances, the proposed full pipeline reduced the word error rate (WER) from 37.98% to 34.36% and the character error rate (CER) from 18.68% to 17.37% compared to the Whisper baseline, achieving the best WER and CER. Furthermore, the proposed method produced only seven high-edit cases and achieved the lowest mean severity in the targeted manual annotation.
더 많은 논문을 준비 중입니다
휴먼교육센터 AI연구소와 휴먼딥브릿지는 직업훈련·교육 AI 분야의 연구를 지속적으로 수행하고 있습니다. 게재 확정 시 이 페이지에 순차 업데이트됩니다.
