Pyler, a video-understanding artificial intelligence company, said two of its research papers have been accepted at a leading global AI conference, covering technology that analyzes emotions carried in the human voice and reduces the impact of faulty training data.
The company said on the 22nd that two papers from its AI research team were accepted at NeurIPS 2026. Both will be presented in Sydney, Australia, on Dec. 10.
The first paper covers CoAT, a technology that lets AI analyze emotion, intonation and tremor in a voice without converting speech into text. It was designed to analyze sound information directly, addressing the limits of text-centered AI, which struggles to detect differences in emotion conveyed by tone of voice. The approach draws on specialized AI models across five areas, including speech content, emotion and pitch. Applied to three speech AI models, it improved 38 of 42 performance metrics, and the emotion recognition accuracy of the Qwen2.5-Omni model rose to about 79% from 55%.
The second paper deals with GAPO, a technology that allows AI to train stably even on data containing errors. It compares evaluation results from the model being trained against a reference model given slight modifications, then reduces the training weight of unstable data whose assessment shifts easily. Applied to four AI models, it generally outperformed eight existing training methods. It also delivered steadier performance than existing error-handling techniques in an experiment in which 20% of the data was deliberately mislabeled.
Pyler plans to apply both technologies to AI that judges the risk level of video content and its suitability for advertising brands. The idea is to use them in reviewing content by analyzing not only scenes in a video but also the speech patterns, emotions and background sounds within it.
"This research is meaningful in that it enables AI to understand even subtle information that people naturally distinguish by ear but find hard to describe in writing," said Kim Jong-seok, head of AI at Pyler. "We will continue to advance the technology so that such information is analyzed without being missed."