基于深度学习的数字音乐流派分类研究  

Deep Learning-Based Digital Music Genre Classification Research

作　　者：唐和铭 游福成[1] 刘玉琴[1] TANG Heming;YOU Fucheng;LIU Yuqin(Beijing Institute of Graphic Communication,Beijing 102600,China)

机构地区：[1]北京印刷学院,北京102600

出　　处：《北京印刷学院学报》2023年第6期37-44,共8页Journal of Beijing Institute of Graphic Communication

基　　金：国家数字复合出版系统工程应用示范(中共宣传部)项目(11000301118)。

摘　　要：针对目前基于深度学习的数字音乐流派分类算法准确性较低的问题,提出了一种基于ECAPA-TDNN的ECAPA-TDNN-BLSTM模型。该模型将双向LSTM网络融合到ECAPA-TDNN模型中,以获取更多关于时间上下文的信息,从而提高模型的准确性。本文使用包含十种音乐流派的1000首歌曲的数据集进行了测试,并在测试集中获得了良好的结果。经过1000个epoch的训练,最后50个epoch的平均准确率为96.07%,证明该模型的有效性。结果表明,与现有模型相比,本文提出的模型对音乐流派的分类识别准确率有一定的提升。To address the problem of low accuracy of current deep learning-based digital music genre classification algorithms,an ECAPA-TDNN-BLSTM model based on ECAPA-TDNN is proposed in this paper.The model fuses a bidirectional LSTM network into the ECAPA-TDNN model to obtain more information about the temporal context and thus improve the accuracy of the model.Meanwhile,this paper uses a dataset containing 1000 music tracks from ten music genres for segmentation tests and obtains good results in the test set.After training with 1000 epochs,the average accuracy of the last 50 epochs is 96.07%,which proves the effectiveness of the model.The results show that the model proposed in this paper has improved the classification and recognition accuracy of music genres compared with the existing models.

关 键 词：时延神经网络 长短期记忆网络 深度学习 数字音乐流派分类 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]