用於陪伴型機器人之輕量化深度學習音樂情緒辨識模型

Abstract

為了應對現今社會高齡化，導致老人缺乏陪伴導致的孤獨問題，本研究提出用於陪伴型機器人Zenbo Junior II的音樂情緒辨識模型來解決老人孤獨導致的情緒問題。在音樂情緒辨識這個研究領域中，雖然也有很多人已經在進行這項研究，但是這些研究中沒有能用於Zenbo Junior II的輕量化架構。本研究提出的方法是使用一維卷機神經網路(1D-Convolutional Neural Network, 1D-CNN)替換掉常用的2D-CNN並且使用閘門循環單元(Gated Recurrent Unit, GRU)使模型能更好的考慮音頻特徵的連續性。在訓練完模型後儲存並應用於Zenbo Junior II上，先將另一研究的情緒對應成4種情緒後播放音樂調適情緒。本研究提出之模型在PMEmo數據集上Valence和Arousal分別為0.04和0.038與其他模型相比效能最好。並且參數量僅有0.721M浮點運算次數僅有9.303M，遠小於其他相比較之模型。運算強度最靠近Zenbo Junior II之最佳工作點，且模型辨識音樂所需推理時間僅需229毫秒，可以即時辨識出音樂的情緒。這些表明本研究成功提出一個輕量化且效能優異，並且可以在Zenbo Junior II上運行的模型。To address the loneliness of the elderly in an aging society, this study proposes a music emotion recognition model for the companion robot Zenbo Junior II. Although many researchers have studied music emotion recognition, no lightweight framework exists for Zenbo Junior II. This study uses a 1D-Convolutional Neural Network (1D-CNN) instead of the commonly used 2D-CNN and incorporates a Gated Recurrent Unit (GRU) to better capture audio feature continuity. After training, the model was saved and deployed on Zenbo Junior II, where emotions from another study were mapped to four categories, and music was played to adjust emotions. The proposed model achieved the best performance on the PMEmo dataset with Valence and Arousal scores of 0.04 and 0.038, respectively. It has only 0.721M parameters and 9.303M FLOPs, significantly smaller than other models. The computational strength is closest to Zenbo Junior II's optimal operating point, and the model's inference time for music recognition is only 229 milliseconds. These results demonstrate that this study has successfully developed a lightweight and efficient model suitable for Zenbo Junior II.