@techreport{oai:ipsj.ixsq.nii.ac.jp:00234669,
 author = {Haitong, Sun and Jaehyun, Choi and Nobuaki, Minematsu and Daisuke, Saito and Haitong, Sun and Jaehyun, Choi and Nobuaki, Minematsu and Daisuke, Saito},
 issue = {57},
 month = {Jun},
 note = {In media technology, comparison of a sequential data with another is a fundamental technique for many applications, and dynamic time warping is often conducted. Conventionally, two sequences of raw features were compared. These days, more abstract and sequential representations are used, which are obtained with deep neural networks. One of these representations is posteriorgram, where each frame is composed of n-dimensional class posteriors, and frame-to-frame distance is often calculated using a divergence metric such as Bhattacharyya distance. In this study, a novel method is proposed to distill the class-to-class distances encoded in any classifier used to calculate posteriors, and the distances are effectively used to accelerate posteriorgram-based DTW by approximating it as DTW between two sequences of most likely classes. Utterance comparison experiments showed that the proposed method can accelerate the distance calculation step in posteriorgram-based DTW by a factor of 30., In media technology, comparison of a sequential data with another is a fundamental technique for many applications, and dynamic time warping is often conducted. Conventionally, two sequences of raw features were compared. These days, more abstract and sequential representations are used, which are obtained with deep neural networks. One of these representations is posteriorgram, where each frame is composed of n-dimensional class posteriors, and frame-to-frame distance is often calculated using a divergence metric such as Bhattacharyya distance. In this study, a novel method is proposed to distill the class-to-class distances encoded in any classifier used to calculate posteriors, and the distances are effectively used to accelerate posteriorgram-based DTW by approximating it as DTW between two sequences of most likely classes. Utterance comparison experiments showed that the proposed method can accelerate the distance calculation step in posteriorgram-based DTW by a factor of 30.},
 title = {Distilling the Class-to-class Distances Encoded in a Classifier to Accelerate DTW of its Posteriorgram},
 year = {2024}
}