The Comparison of Different Deep Learning Models for Chinese Musical Instrument Recognition

Instrument recognition plays a crucial role in various domains such as music, education, artificial intelligence, and cultural preservation. While deep learning technologies have achieved significant success in the recognition of Western musical instruments, their application to Chinese musical instruments remains relatively unexplored. To bridge this gap, this study investigates the use of deep learning techniques for the recognition of traditional Chinese instruments. Four deep learning models—RNN, ResNetSE, TDNN, and ECAPA-TDNN—were chosen as experimental models for the task. A diverse dataset consisting of 37 traditional Chinese instruments, including the Erhu, Pipa, Guzheng, and Dizi, was created. The dataset incorporates different tones, timbres, and playing speeds, ensuring a comprehensive representation of Chinese musical instruments. For feature extraction, Mel-frequency cepstral coefficients (MFCC) and spectrograms were employed to capture significant audio features from the instrument sounds. The Chinese instrument recognition system processes the audio through three stages: preprocessing, feature extraction, and deep learning-based classification. The study further compares the performance of MFCC and spectrograms in terms of recognition accuracy. Experimental results demonstrate that MFCC provides superior feature extraction results. Among the models using MFCC features, ECAPA-TDNN achieved the highest accuracy at 97.18%, followed by TDNN at 96.02%, ResNetSE at 95.01%, and RNN at 92.59%. These findings suggest that the TDNN-based neural networks are particularly effective for recognizing Chinese instruments.