融合梅尔谱和循环残差的小样本音频分类模型

针对小样本环境下音频信号分类精度急需提高的问题,首先提出自适应梅尔滤波算法提取具有更高区分度的梅尔谱图,再提出循环残差结构并结合迁移和微调构建循环残差网络频谱分类器,融合自适应梅尔滤波算法和循环残差网络频谱分类器生成一种主要用于小样本环境的音频信号分类模型。 以 ESC- 50、music speech、Free ST Chinese Mandarin Corpus(FSCMC)为源数据集模拟四个不同属性的小样本环境。 仿真显示在各小样本环境下生成模型的分类精度与 MF-VGG16、 10 layers CNN、CRBM 等模型相比均有一定程度的提高,且精度曲线更平滑,性能更稳定。In view of the urgent need to improve the classification accuracy of audio signals in small sample environment, an adaptive Mel filter algorithm was proposed to extract Mel spectrum with higher discrimination, and then a cyclic residual structure was proposed. Combined with migration and fine tuning, a cyclic residual network spectrum classifier was constructed, An audio signal classification model mainly used in small sample environment was generated by combining adaptive Mel filter algorithm and cyclic residual network spectrum classifier. Esc-50,music speech and Free ST Chinese Mandarin Corpus (FSCMC) were used as source data sets to simulate four small sample environments with different attributes. The simulation shows that the classification accuracy of the generated model in each small sample environment is improved to a certain extent compared with MF-VGG16,10 layers CNN,CRBM and other models, and the accuracy curve is smoother and the performance is more stable.