古筝声音品质主观评价指标的CRNN量化方法

CRNN Quantitative Method for Subjective Evaluation Index of Guzheng’s Acoustic Quality

古筝等民族乐器声音品质好坏往往采用多位专家共同参与打分的方法来评定。这种评价主观性强、偏差大,且效率低、成本高,因此提出一种基于深度学习的古筝声音品质主观评价指标量化方法。在确定总体评价古筝声音品质的五个主观指标之后,该方法选择卷积循环神经网络(CRNN)作为深度学习模型框架,将实验收集到的古筝音频信息作为输入,以专家对各指标的综合评价结果作为模型监督,训练深度学习的主观评价指标量化模型。经过检验,这种方法达到了专家主观评价的综合效果,满足了实际检验需求。

As a Chinese musical instrument,the acoustic quality of Guzheng is mainly evaluated by a group of experts.This method is very subjective,costly,less effective and lack of objectivity and accuracy.This paper proposes a quantitative method of subjective evaluation index of acoustic quality based on deep learning.First of all,five indicators for Guzheng’s acoustic quality evaluation are determined.Then,the convolution recurrent neural network(CRNN)is chosen as the model frame for deep learning,the Guzheng’s audio information collected in the experiment is selected as the input,and the evaluation results of experts on each indicator are introduced as a model supervision.Finally,the deep learning model for quantitative evaluation index based on CRNN is trained.The results show that this model algorithm is basically consistent with the comprehensive evaluation from experts’subjective assessments,and can meet the actual inspection requirements.